DEV Community: Furkan Gözükara

Install ComfyUI and Learn FLUX 3 Klein: The Complete Beginner Tutorial

Furkan Gözükara — Mon, 19 Oct 2026 02:39:07 +0000

Install ComfyUI and Learn FLUX 3 Klein: The Complete Beginner Tutorial

Install ComfyUI on Windows and make precise FLUX 3 Klein images: 16 core examples, RTX 5090 tests and launcher tuning. The package includes RunPod, SimplePod, Massed Compute and Linux routes.

Links
Info
Chapters
18 October 2026 Update V143
17 October 2026 Update V142
10 October 2026 Update V140

Info

You will learn installation, models, captions, image edits, layouts, batches, shared storage, launch arguments and updates. Follow the complete route from the download post to a working image. Open saved workflows, place references, inspect enlarged results and compare changes using sliders and labeled boards. The examples show exact lettering, object preservation and ten-reference composition.
Main topics include HQ INT8 ConvRot, Gemini, seeds, LoRAs and VRAM. Benchmarks: up to 2.3x on RTX 3090 and 1.82–2.18x on RTX 5090, for warm transformer steps only; loading, text encoding and image decoding excluded. Learn what --reserve-vram changes, why a saved BAT needs a server restart, and how to verify a small workflow afterward. Change one option at a time and compare the same task.
Use the chapters to jump to each setup step, feature and comparison.

Chapters

Time	Chapter
00:00:00	Precise edits and 16 FLUX 3 examples
00:02:47	Open a preset and generate your first image
00:04:00	Package and Windows requirements
00:04:54	Download and extract the installer
00:06:26	Install ComfyUI and Python
00:07:11	Finish Torch and CUDA setup
00:07:56	Install custom nodes with bundle 100
00:09:08	Download the model downloader
00:10:02	Start and explore the downloader
00:11:19	Set model folders and ComfyUI structure
00:11:45	Select the FLUX 3 core bundle
00:12:36	Download and verify three models
00:13:06	Check model filenames and folders
00:13:56	Use the URL Downloader
00:14:27	Start ComfyUI and explore nodes
00:15:20	Open the advanced image preset
00:16:03	Describe the aurora and inspect loaders
00:16:41	Set shape, megapixels and dimensions
00:17:17	Seed, sampler, steps and guidance
00:17:46	Generate and inspect the aurora
00:19:05	Create a crystal kingfisher
00:19:58	Generate exact poster lettering
00:21:06	HQ INT8 ConvRot conversion
00:21:21	RTX 3090 speed and mixed precision
00:22:34	RTX 5090: BF16 versus HQ INT8
00:23:44	RTX 5090 live generation
00:25:08	Give the prompt guide to your LLM
00:26:18	Use Gemini for captions
00:27:15	Expand scene, poster and edit ideas
00:28:30	Generate from the enhanced caption
00:29:38	Open image editing and load a reference
00:30:22	Describe changes and preserved details
00:31:29	Inspect the before-and-after slider
00:32:16	Remove, move and replace desk objects
00:34:03	Reinterpret a painting as photography
00:35:15	Keep cat identity in a new setting
00:36:32	Combine identity and style references
00:38:22	Transform a landscape into winter
00:39:23	Compare geography and seasonal light
00:39:54	Replace packaging lettering
00:40:54	Inspect the enlarged tea label
00:41:25	Assign reference roles
00:42:37	Generate the combined campaign image
00:43:30	Layout boxes and Modern Node Design
00:44:35	Draw four placement regions
00:45:32	Match object mentions to box labels
00:46:16	Inspect all four climates and labels
00:47:12	Load ten references in order
00:48:37	Set reference resolution and caption
00:49:23	Check all ten cabinet objects
00:50:20	Create three product views
00:51:27	Images per prompt and batch results
00:52:46	Prepare a folder of text prompts
00:53:52	Queue and inspect both outputs
00:54:36	Compare prompt weighting
00:55:37	Seeded choices, schedules and BREAK
00:56:47	Negative prompts and CFG 1
00:57:18	Repeat a fixed seed
00:58:21	Randomize the composition
00:58:55	Variation seed and strength
01:00:08	Configure a compatible LoRA
01:01:07	Reuse models with extra paths YAML
01:01:50	Move the decoder and save its mapping
01:03:13	Restart and verify shared models
01:03:41	Edit Default BAT launcher arguments
01:04:27	Add `--reserve-vram` 2
01:05:34	Restart and verify the changed launcher
01:06:15	Cache and low-VRAM options
01:07:13	Remove or restore an argument
01:08:31	Optimized launcher and memory menu
01:09:31	Choose caching and verify launch mode
01:10:51	Update while preserving your work
01:11:33	Update ComfyUI and the node bundle
01:12:52	Test the updated setup
01:14:21	Continue with your reusable installation

For anyone building a reusable ComfyUI setup. Get help through the pinned comment, Patreon, its Discord link or the comments.

18 October 2026 Update V143

17 October 2026 Massively Detailed Windows Tutorial : https://youtu.be/yOj9PYq3XYM

Presets Folder Content In Latest Zip File

This is a MOSS-Image update. V143 adds six new presets for MOSS-Image, the new open text-to-image model from OpenMOSS: Text To Image with the official 50-step quality recipe, Fast 25 Steps, Turbo 8 Steps, Layout Boxes, and Base model versions of Text To Image and Layout Boxes. It also adds a MOSS-Image prompt guide for ChatGPT, Claude and Gemini, a ComfyUI notebook for Google Colab, an updated AvatarForever preset with a live preview panel, fixes for the VRAM-optimized launcher and the installer, the newest ComfyUI with native MOSS-Image support, and the matching custom-node and SwarmUI extension updates.

Four native 1920×1088 images from the new Advanced Text To Image preset, with no upscaling: 50 steps, 32 to 41 seconds each on an RTX 5090. Every label and poster line came out exactly as written, and the chalkboard is spelled correctly.

Every result in this post comes from a real run of the new presets with their default settings; only the prompt (for the layout, also the boxes), seed and output size were changed. The timings were measured on our RTX 5090 after the models were loaded.

Here is the complete Presets folder of V143: 96 ready-to-use ComfyUI presets, plus 3 API presets in the api_paid folder. The six new MOSS-Image presets are marked NEW and the updated AvatarForever preset is marked UPDATED:

The presets cover image, video, lip sync, audio and music models: AvatarForever, ERNIE Image, FLUX.1 Dev, Kontext and SRPO, FLUX 2, FLUX 3 Klein, Ideogram 4, Krea 2, LTX 2, LTX 2.3, LTX 2.5, MiniMax H3, MiniMax Music 3, MOSS-Image, Phantom Wan, Qwen Image, SDXL, Stable Video Infinity 2 Pro, Wan 2.1, Wan 2.2 and Z-Image.

NEW PRESETS — MOSS-Image: Text To Image, Fast, Turbo, Layout Boxes and Base

"MOSS Image - Advanced Text To Image - 261017"
"MOSS Image - Fast 25 Steps - 261017"
"MOSS Image - Turbo 8 Steps - 261017"
"MOSS Image - Advanced Layout Boxes - 261017"
"MOSS Image Base - Advanced Text To Image - 261018"
"MOSS Image Base - Advanced Layout Boxes - 261018"

Here are the new and updated presets, with what each one does:

MOSS-Image is a new open image model from OpenMOSS: a 5B diffusion transformer that reads your prompt with the Qwen3.5-9B language model and uses the FLUX 2 VAE. It renders clear, accurate text inside images, accepts long and detailed prompts of up to about 4,000 tokens, and, as an experimental feature of the model, can place objects, characters and text in boxes you draw. It is released under the Apache 2.0 license, which also allows commercial use. It makes images from text; it does not edit images or use reference images.

SFT and Base: SFT is the official fine-tune of Base on curated data, and the four main presets use it; in the official comparison it groups text more clearly. Base is the raw model without post-training, with more varied images, meant for LoRA training; two Base presets are included.

All six presets share the same easy controls:

One prompt box: plain language or tags, and in the Text To Image presets also a JSON layout. Long, detailed prompts work best. Put exact text for signs, labels and posters in quotes, and use real line breaks to control the lines.
Output size: choose an aspect ratio and megapixels, or type the width and height. Any aspect ratio works; 1 to 2 megapixels is recommended. Sizes snap to multiples of 16.
Batches: + Prompt adds more prompts to one run, a folder of .txt files runs a whole prompt list (JSON layouts too), and Images per prompt makes several images of each prompt.
Prompt syntax: random choices {red|blue} follow the seed in the main prompt, added prompts and folder prompts, so a fixed seed repeats the same prompt. Schedules [red:blue:0.5] and alternation [red|blue] work in plain-text prompts of the Text To Image presets. Weights like (word:1.2) are removed because MOSS-Image's text encoder would read them as words; describe emphasis in words instead.
JSON layouts in the Text To Image presets: a complete MOSS-Image layout is sent unchanged, only without extra spaces. Pixel-based layout drafts from the official MOSS prompt enhancer or our prompt guide are converted to the model's format automatically and scaled to your output size. In the Layout Boxes presets, the prompt box holds only the caption of the whole image.
Automatic SFT de-band: the SFT model can draw faint vertical color stripes on skies, fog and plain walls. The presets apply a community fix to SFT automatically and skip it for Base.
Seed: fixed, or randomized on every run, with an optional variation seed and strength.
Optional LoRA stack for MOSS-Image LoRAs, off by default.
Attention set to PyTorch: it matched the official MOSS-Image images in our tests, and it also prevents black images with the BF16 files when ComfyUI starts with Sage attention, as our launchers do (RTX 3090 test). Keep it on PyTorch in your own BF16 workflows.
The model, text encoder, VAE and latent are already connected. To use the BF16 files, switch the model loader. Use Base only in the 50-step presets: the Fast and Turbo recipes were made for SFT.

Here is the Advanced Text To Image preset after a real run of a magazine cover prompt:

All five text lines came out exactly as written, from the huge WANDER title to the small issue line. The 1024×1536 image took 21.8 seconds with the official 50-step recipe.

Fast 25 Steps, Turbo 8 Steps and Base: here is the same prompt and seed through all four text-to-image presets at 1920×1088:

Advanced Text To Image (SFT, 50 steps, Euler / normal, CFG 3, shift 5, the official recipe): the most faithful results, especially for text and designs. 34.3 seconds here, 12.9 seconds at 1024×1024.
Fast 25 Steps (ER-SDE / simple, CFG 3, Temporal Score Rescaling 0.95, layer skip 22 and 24 from 10% to 70% of the steps): 16.6 seconds here, about half the time of the 50-step preset. Large text holds up well; very small text can occasionally lose a letter.
Turbo 8 Steps (the community 8-step LoRA by lodestones at 0.7, Euler / simple, CFG 1): 3.4 seconds here, about a tenth of the 50-step time. The 8-step LoRA is a work in progress, so its images differ more from the 50-step ones and can show extra signs.
Base Advanced Text To Image (Base model, the official 50-step recipe): more varied images; 32.5 seconds here. For text-heavy designs, start with SFT: in the official Base versus SFT example, SFT groups text more clearly.

Layout Boxes: draw a box for each object, character or line of text and describe it. The main prompt becomes the caption of the whole image:

Each box is an object or exact text, with its own description. To turn object boxes into characters, type their numbers in character_boxes, here 8 for the barista.
The canvas follows the output size, and the preset builds MOSS-Image's layout JSON from your boxes.
The box editor needs Settings → Modern Node Design (Nodes 2.0). Turn it on once.
Boxes guide placement. They are not masks, so positions can vary a little.

Here is the result of that layout in the same preset:

Every box landed where it was drawn: the headline, three drinks with their exact names and prices, and the waving barista as a character. The 1536×1024 image took 22.6 seconds.

MOSS-Image, the 8-step LoRA and the de-band fix are released under the Apache 2.0 license.

UPDATED PRESET — AvatarForever: Live Preview Panel and Tested Mouth-Movement Settings

"AvatarForever Unified Image or Text + Audio To Long Video - 260920"

New Live Preview & Final Video panel to the right of the generation controls. It shows a frame preview after each sampling step, starting with the first, then plays the finished MP4.
New tested preset settings: a 512×768 canvas and 16 latent frames per chunk, instead of 832×1248 and 4. In our singing demo, 4-frame chunks kept the lips closed, while 16-frame chunks gave sustained mouth movement through the whole 18-second song. At 1024×1024 the mouth stayed mostly closed even with 16-frame chunks, so check a short sample when you change the image, audio, resolution or chunk size. Larger chunks need more memory; 4-frame chunks still work for speech and use less.
Sharper previews: the live preview uses the small taeltx2_3 preview decoder when it is installed. It is now in the LTX 2.3 Avatar Forever Core Bundle of our SwarmUI Model Downloader V183. Without it, the preview uses a coarser latent view.
Your own AvatarForever workflows keep working and now show live previews on the sampler. Set live_preview to off to turn them off.

NEW PROMPT GUIDE — MOSS-Image Feed for ChatGPT, Claude and Gemini

A new MOSS_Image_Enchance_Prompt_Feed_For_LLMs.txt is included. Give it to ChatGPT, Claude or Gemini together with your request, and it writes prompts that MOSS-Image understands:

Detailed captions for photos, illustrations and designs, with the exact visible text in quotes and where and how it appears.
Layouts as pixel-based JSON drafts for the Text To Image presets, which convert them to MOSS-Image's own format automatically, or a box-by-box plan to draw in the Layout Boxes editor.
Tag-style prompts for anime, in the recommended tag order.
The right settings: output size, CFG for short prompts, the negative prompt and when to use SFT or Base.

NEW — ComfyUI Colab Master Notebook for Google Colab

"ComfyUI_Colab_Master_Notebook.ipynb"

A new Google Colab notebook runs ComfyUI with ComfyUI-Manager on a Colab GPU and opens it in your browser through a Cloudflare link. It sets up Qwen-Image 2.1 for you:

Install: ComfyUI (the newest stable release, or master if you choose it) with ComfyUI-Manager and cloudflared, which makes the link without an account. Colab's own PyTorch is kept.
Qwen-Image 2.1 in the right size: the official ComfyUI files are picked from your GPU memory and system RAM: INT8 with the W4A8 text encoder (14.2 GB) when either is below 20 GiB, as on a T4; INT8 with the INT8 text encoder (17.3 GB), the official template's files, when both have at least 20 GiB; or full BF16 (32.4 GB) when you choose it. Broken downloads resume when you run the cell again.
Any other model: paste up to four direct download links, such as LoRAs or upscalers, and choose the folder or let the notebook read it from the link.
Start: you get a trycloudflare.com link, and ComfyUI opens with the Qwen-Image 2.1 text-to-image workflow already loaded, with the official settings (Euler, simple, 25 steps, CFG 1, 1024×1024).
Keep your work: one cell packs your outputs and saved workflows into a ZIP to download. The notebook also explains how to upload images and videos, and what is kept when the runtime ends.

Open Google Colab, choose Upload notebook, select the file, pick a GPU runtime (T4 or better) and run steps 1 to 4 in order. T4 GPUs can be allocated on free accounts, but availability is not guaranteed, and Colab says free runtimes used mainly through an external web UI may be terminated. The Cloudflare link is public, so don't share it. The Qwen-Image 2.1 weights use the Qwen Research License, which is non-commercial.

LAUNCHER AND INSTALLER FIXES

Windows_Run_VRAM_Optimized.bat: option 3 (normal VRAM), and VRAM choice 3 in Custom mode, passed --normalvram, which ComfyUI removed in version 0.22.0, so ComfyUI stopped at once with "unrecognized arguments: --normalvram". Both now start ComfyUI with its default memory management and no extra flag.
Windows_Run_VRAM_Optimized.bat also describes the memory modes as current ComfyUI uses them. LOWVRAM's --lowvram has no effect while dynamic VRAM is on, the default on NVIDIA GPUs, and otherwise only runs the text encoders on the CPU; NOVRAM turns dynamic VRAM off. Pressing Enter in the cache menu now keeps ComfyUI's own default, RAM-pressure caching, instead of the older classic caching; choose 1 for classic caching.
Windows_Install_Or_Update_ComfyUI.bat: the pip update and the uv install now ignore your personal pip settings (the pip.ini in your user profile and PIP_ environment variables), so a custom package index set there for another project can no longer stop the installer at this step. Its Windows requirements tutorial link now points to the current video: https://youtu.be/ptuLL0C72Ms

HOW TO UPDATE

Download the latest ComfyUI_V143.zip from this post.
Extract it over your existing installer folder and overwrite the older files.
Close ComfyUI, then run Windows_Install_Or_Update_ComfyUI.bat. It updates ComfyUI to the newest version, with MOSS-Image support. The MOSS-Image presets do not run on older ComfyUI.
Run Windows_Custom_Nodes_Bundles_Installer.bat and install/update bundle 100. It updates SECoursesAudioTools with the new MOSS-Image nodes and the AvatarForever Live Preview & Final Video node that the updated AvatarForever preset uses, and keeps FoleyExtension, SwarmUI ExtraNodes and ComfyUI-Easy-Use, which the MOSS-Image presets also use, up to date.
Restart ComfyUI and refresh the browser tab.
For the Layout Boxes editor, turn on Settings → Modern Node Design (Nodes 2.0).

Use our latest SwarmUI Model Downloader (SwarmUI_Model_Downloader_v183.zip or newer) and download the MOSS-Image Core Bundle, the first card of the SwarmUI Bundles tab. It has the six files the six presets use, about 23.2 GB in total (the downloader counts in GiB, so it shows 21.64 GB):

moss_image_sft_int8_convrot.safetensors (5.41 GB), the SFT model in INT8 ConvRot.
moss_image_base_int8_convrot.safetensors (5.41 GB), the Base model in INT8 ConvRot.
qwen3.5_9b_int8_convrot.safetensors (9.75 GB), the Qwen3.5-9B text encoder in INT8 ConvRot.
Flux/flux2-vae.safetensors (0.34 GB), the FLUX 2 VAE, the same file the FLUX 2 and FLUX 3 presets use.
MOSS_Image_SFT_DeBand_L37-41.safetensors (1.13 GB), the SFT de-band fix.
MOSS_Image_SFT_8_Steps_DMD_LoRA.safetensors (1.19 GB), the 8-step LoRA of the Turbo preset.

The INT8 ConvRot model files are about half the size of BF16, and in our RTX 3090 test they were about twice as fast (48 versus 99 seconds at 50 steps). The BF16 models, a smaller W4A8 text encoder (select it in the text encoder loader), the small-decoder FLUX 2 VAE and the extra de-band fix for layers 32–36 are optional downloads under Image Generation Models → MOSS-Image Models. The de-band author recommends the extra fix for images 1536 pixels wide or larger and for big, smooth skies; select it in the de-band node.

SwarmUI Installer and Unified Massive Models Downloader: https://www.patreon.com/posts/114517862

17 October 2026 Update V142

This is a FLUX 3 Klein update. V141 adds three new presets for Black Forest Labs' new FLUX 3 Image Klein model: Text To Image, Image Edit with up to 10 reference images, and Layout Boxes. It also adds a FLUX 3 Klein prompt guide for ChatGPT, Claude and Gemini, our new HQ INT8 ConvRot FLUX 3 Klein model, ComfyUI 0.41.0 with native FLUX 3 support, and the matching custom-node and SwarmUI extension updates.

A native 3840×2160 image from the new Text To Image preset, with no upscaling: 16 steps in 23 seconds on an RTX 5090.

Every result in this post comes from a real run of the new presets with their default 16 steps, Euler / simple, CFG 1. The timings were measured on our RTX 5090 after the models were loaded.

NEW PRESETS — FLUX 3 Klein: Text To Image, Image Edit and Layout Boxes

"FLUX 3 Klein - Advanced Text To Image - 261015"

"FLUX 3 Klein - Advanced Image Edit - 261015"

"FLUX 3 Klein - Advanced Layout Boxes - 261015"

FLUX 3 Image Klein is a 5.6B image model with the Qwen3-VL-4B text encoder. One model does text to image, image editing, up to 10 reference images, layout boxes and text rendering. Our presets run the fast distilled model with its recommended 16 steps and no guidance (CFG 1), using Euler / simple.

All three presets share the same easy controls:

One prompt box with an image gallery: write the whole scene. Add up to 10 images and refer to them as @image1 … @image10 (@img1, @picture1 and @pic1 also work). With no image, it is text to image.
Output size: choose an aspect ratio and megapixels, or type the width and height. Sizes snap to multiples of 16.
Batches: + Prompt adds more prompts to one run, a folder of .txt files runs a whole prompt list, and Images per prompt makes several images of each prompt.
Advanced prompt syntax: weights (ceramic:1.1), random choices {red|blue}, schedules [red:blue:0.5], alternation [red|blue] and <break>. Random choices follow the seed, so a fixed seed repeats the same prompt.
Seed: fixed, or randomized on every run, with an optional variation seed and strength.
Reference size: one Reference megapixels setting resizes every reference image and keeps its shape (1 MP by default, up to 4 MP). It is separate from the output size.
Optional LoRA stack for FLUX 3 LoRAs, off by default.
The model, text encoder, VAE and latent are already connected. There is nothing else to set up.

Here is the Text To Image preset after a real run of a poster prompt with English and Turkish text:

All four text phrases came out exactly as written, including the Turkish characters. The 1024×1536 image took 3.2 seconds. The first run takes longer because it loads the models.

Image Edit: the same workflow with your image in the gallery. Here it turns a celadon museum vase cobalt blue and keeps its shape, handles, framing and shadow:

Say what changes and what stays, and refer to your photo as @image1.
Match the output aspect ratio to your photo to keep its framing. This 1024×1536 edit took 5.2 seconds.
The preset opens with a sample card. Add your own image with + Add reference images and remove any card marked "file missing".

Up to 10 reference images in one prompt: here ten separate museum photos became one display wall, each object in its requested niche:

All ten objects appear in the requested order and closely match their photos. The 1680×944 image took 41 seconds with all ten references at 1 MP each. The reference photos are public domain (CC0) images from The Cleveland Museum of Art.

Layout Boxes: draw a box for each object or line of text, describe it, and refer to it in the caption as <obj_1>, <obj_2> and so on:

Each box is an object or exact text, with its own description. The preset builds FLUX 3's layout table from your boxes, and the boxes follow the output size.
The box editor needs Settings → Modern Node Design (Nodes 2.0). Turn it on once.
Boxes guide placement. They are not masks, so positions can vary a little.

Here is the result of that layout in the same preset:

All four terrariums and captions landed where their boxes are, and the four captions are spelled exactly. The 1024×1024 image took 2.5 seconds.

FLUX 3 Klein is released under the FLUX Kommunity License: free for non-commercial use, and people and companies with a yearly revenue under US$5 million (affiliates included) may use the images commercially under its conditions. Read the license for the full terms.

NEW MODEL — Our FLUX 3 Klein HQ INT8 ConvRot

"FLUX-3-Image-Klein-INT8-ConvRot-HQ.safetensors"

The presets use our own INT8 ConvRot conversion of the distilled FLUX 3 Klein model. On the updated ComfyUI it loads with the normal Load Diffusion Model node, so no special loader is needed.

6.07 GB instead of 11.1 GB for the original BF16 model. While denoising at up to 1024×1024, the model alone peaked at about 6 GiB of VRAM instead of about 10.6 GiB for BF16 (RTX 3090 test).
About 2× faster per step than BF16 (1.9–2.3× in our RTX 3090 tests, model only).
Closer to BF16 than the official INT8 file: in our test of 40 sampling states, its mean output error against BF16 was 2.14%, versus 3.05% for the official INT8 ConvRot file with its default settings and 2.27% with matched attention. One held-out editing case favored the official file with matched attention. In exchange, our file is 0.25 GB larger and was 6–22% slower per step in the same RTX 3090 test. These are numerical measurements made with the BF16 text encoder, not image-quality percentages.
The first text block and the last joint block, the most sensitive in our tests, stay in BF16, as do the small conditioning, normalization and output layers. The other 164 linear layers use native INT8 ConvRot.

NEW PROMPT GUIDE — FLUX 3 Klein Feed for ChatGPT, Claude and Gemini

A new FLUX_3_Klein_Enchance_Prompt_Feed_For_LLMs.txt is included. Give it to ChatGPT, Claude or Gemini together with your request and images, and it writes prompts that FLUX 3 Klein understands:

Complete scene captions for text to image, with the most important details first.
Edit prompts that say what changes and what must stay, and multi-reference prompts that give every image a clear role.
Exact text for posters and designs, in the original language and spelling.
Layout prompts in the FLUX 3 box format, with the <obj_1> labels these presets use.
The right settings and the ComfyUI / SwarmUI details, such as @image1 meaning FLUX 3's first reference image.

AvatarForever Preset Now Has Live Preview Feature

2:3 and 3:2 aspect ratios works best - 1:1 not that good

HOW TO UPDATE

Download the latest ComfyUI_V141.zip from this post.
Extract it over your existing installer folder and overwrite the older files.
Close ComfyUI, then run Windows_Install_Or_Update_ComfyUI.bat. It updates ComfyUI to the newest version, with FLUX 3 support. The FLUX 3 presets do not run on older ComfyUI.
Run Windows_Custom_Nodes_Bundles_Installer.bat and install/update bundle 100. It updates SECoursesAudioTools with the new FLUX 3 nodes and keeps FoleyExtension, SwarmUI ExtraNodes and ComfyUI-Easy-Use, which the presets also use, up to date.
Restart ComfyUI and refresh the browser tab.
For the Layout Boxes editor, turn on Settings → Modern Node Design (Nodes 2.0).

Use our latest SwarmUI Model Downloader (SwarmUI_Model_Downloader_v181.zip or newer) and download the FLUX 3 Klein Core Bundle, the first card of the SwarmUI Bundles tab. It has the three files all three presets use, about 11.7 GB in total (the downloader counts in GiB, so it shows 10.85 GB):

FLUX-3-Image-Klein-INT8-ConvRot-HQ.safetensors (6.07 GB), our HQ INT8 ConvRot model.
qwen3vl_4b.safetensors (5.24 GB), the FP8 Qwen3-VL-4B text encoder.
Flux/flux2-vae.safetensors (0.34 GB), the FLUX 2 VAE, which FLUX 3 uses too.

The text encoder is the same file Krea 2 uses and the VAE is the FLUX 2 VAE, so if you already have those, only the 6.07 GB model is new.

SwarmUI Installer and Unified Massive Models Downloader: https://www.patreon.com/posts/114517862

10 October 2026 Update V140

ComfyUI V140 is ready. Here is everything we changed since V138: 4 new MiniMax H3 presets, optional MiniMax H3 RefMods (reusable reference files) in our MiniMax H3 presets, a native sparse attention engine in our MiniMax H3 Speed Optimizer, fixes that keep our nodes working on the newest ComfyUI (tested on ComfyUI 0.39.0), and fixes for every preset that failed when we ran all presets with real generations.

New MiniMax H3 Presets

MiniMax H3 Lip Synch - Turbo 4 Steps (672x1216 and 832x1248)

Here is the 672x1216 preset after a real run with its demo audio and portrait:

Fast talking videos from your audio: upload an audio file, optionally add an image, describe the scene and run. The video length follows your audio. Without an image it works as audio + text. With a video in the reference gallery, it makes a new performance from that video.
Turbo 4 steps by default with the LightX2V FL2VA Turbo 4-step v1.2 LoRA, at 24 fps. A 15-second video took about 165 seconds at 672x1216 on an RTX 5090. Set "4. QUALITY / SPEED / VRAM" to FULL QUALITY for the 50-step path.
Two portrait resolutions: 672x1216, or 832x1248 for more detail. These are 4-step portrait versions of our existing MiniMax H3 Lip Synch preset (1024x1024, 8 steps).

The 832x1248 preset saves the original video first, then runs the new CodeFormer mouth pass:

The 832x1248 preset adds a CodeFormer mouth pass (on by default). It restores only the mouth area of the finished video and saves it as a second MP4, next to the original video.
Also included: our optional face refinement (off by default) and the new optional RefMods.
Tip: for your own audio, put its exact words in the <d>[English] ... </d> part of the scene prompt, or clear the scene prompt to use the automatic one. 5-15 seconds per generation is recommended; longer audio still works in one pass.

MiniMax H3 Long Lip Synch - Native Latent Continuation - Full Audio

Here is the Long Lip Synch preset with its default settings:

One full soundtrack in, one continuous video out: give it a complete audio file plus a portrait image or gallery references. It makes one continuous 24 fps lip-synced video as long as your audio. There is no duration setting and no cutting.
How it stays continuous: it generates overlapping windows of about 10 seconds (243 frames) and carries the last 1.6 seconds (39 frames) of each window forward as clean latents, so nothing is re-encoded between windows. The audio is encoded once, and your original soundtrack is put back into the final MP4.
Works with references and RefMods in every window, unlike the TaoMate streaming preset. The first-frame image anchors only the start.
Defaults: 832x1248 and Turbo 4 steps (FL2VA 4-step v1.2, or Ref2VA 4-step v0.1 when you use references). Use 512x768 for faster drafts. SOL sparse attention and FirstBlockCache are separate switches, both off by default.
Outputs: the final MP4 with your original audio, plus a JSON report with the timing of every window. On an RTX 5090, 30 seconds of 512x768 video took 105 seconds with normal attention.
Experimental: very long videos can change framing over time. In our 15-clip test, 13 clips passed; the 45 and 60-second clips at 832x1248 drifted into close-ups.

MiniMax H3 RefMod Builder - Images, Video, Audio

Add your source media, name your RefMod and run. Only the MiniMax H3 VAEs are used:

Build your own RefMod in one run: add photos of one subject, a video (its soundtrack is included), audio, or a mix. Name it and run. Only the MiniMax H3 VAEs run: no diffusion model, no text encoder and no training.
Options: source resolution (512 by default), "one character" (stacks the photos into one reference) or "separate references" (each photo gets its own picture label), and optional compression (32, 16 or 8) for fewer tokens at the cost of detail.
Ready to use: files are saved to ComfyUI/models/refmods and existing files are never overwritten. Pick the file in any MiniMax H3 preset's Optional RefMods node and write @refmod1 in your prompt.

MiniMax H3 RefMods: Reusable References

A RefMod is a small .safetensors file that stores already-encoded MiniMax H3 references: images, video and/or audio of a character, scene or voice. Build it once, then reuse or share it in any MiniMax H3 preset without encoding the source media again. It is not a LoRA and needs no training.

Here is the new Optional RefMods node with an imported RefMod, used as @refmod1 in the prompt:

New "MiniMax H3 Optional RefMods" node in all 9 existing MiniMax H3 generation presets (Text To Video, Image or Video To Video, References To Video, Audio Only, Text With References To Image and Lip Synch) and in the new presets. It is empty by default, so your results do not change unless you add a RefMod.
Easy picker: "+ Add RefMod", "Import files" and "Refresh" buttons. Each row has an on/off box, the file (with its contents, for example "image + audio"), a strength from 0 to 1, a visual/audio filter and up/down ordering. Imported files go to ComfyUI/models/refmods and are checked before use.
Use them in your prompt with @refmod1, @refmod2 and so on. They are numbered after your gallery images, videos and audio.
Strength sets how much reference detail is kept. It is not a LoRA weight.
RefMods alone are enough: the reference mode and auto mode nodes count RefMods as references, so a RefMod without any gallery image switches generation to the references (Ref2VA) path.
Compatible files: RefMods made with ComfyUI-MiniMaxH3Mod and the Fantastic MiniMaxH3 PromptBuilder load too, without installing those node packs. Their MIT-licensed format code is credited in our notices.
TaoMate streaming does not use RefMods; its preset note now points you to the new Long Lip Synch preset for RefMods with long audio.

MiniMax H3 Speed Optimizer: Native Sparse Attention

Here is the MiniMax H3 Speed Optimizer node, shown in the Text With References To Image preset, where it is off by default:

SOL sparse attention now runs on ComfyUI's own MiniMax H3 sparse attention (our ComfyUI-TeaCache fork). The old bundled attention kernels are gone, so there is nothing extra to compile.
Faster than our old SOL: the old version repeated work and was slower than normal attention at 832x1248 (207 s against 116 s). With the new default protected SOL settings, a 10-second 832x1248 Turbo 4-step video took 99 seconds instead of 111 seconds with normal attention (1.12x, RTX 5090, FirstBlockCache off for this test). At 512x768 the gain was 1.06x. Sparse attention and caching are approximations, so check the quality of your own results.
Auto mode measures before it switches: it uses sparse attention only when it is at least 5% faster. The Text To Video and Image or Video To Video presets have the optimizer on in auto mode, so they now use the native path automatically when it pays off.
FirstBlockCache works together with other speed-ups: it now keeps patches that are already on the model, such as the sparse attention added by the lip-sync Turbo node, instead of replacing them.
Old presets load unchanged. The main switch is now called "H3 optimizations" instead of "4x speedup", and "native" is the new default backend.

Other MiniMax H3 Improvements

Correct colours in streamed MiniMax H3 videos: TaoMate streaming and long lip-sync videos are now tagged with the colour format they are encoded in (BT.601 limited range), so players read the colours correctly.
No torchaudio needed for MiniMax H3 audio: init audio, streaming and long lip sync now resample any soundtrack with ComfyUI's own resampler, with bit-identical results (details below).

Works With the Latest ComfyUI

ComfyUI removed torchaudio and changed its attention and Linear layer code. We updated our nodes so everything keeps working on the newest ComfyUI:

No more torchaudio needed: our MiniMax H3 streaming and long lip-sync nodes (SECoursesAudioTools), the init audio nodes (FoleyExtension) and the audio models in our ComfyUI-WanVideoWrapper fork (HuMo, Ovi, FantasyTalking, MultiTalk and LongCat) now resample audio with ComfyUI's own resampler. Before, on installs without torchaudio, they failed on any soundtrack that was not already at the model's sample rate, and HuMo and Ovi did not load at all. The output is bit-identical: 88 test comparisons had zero difference, and a real MiniMax H3 Turbo 4 video was pixel-identical on all 120 frames.

Here are the new torchaudio-free Mel-Band RoFormer nodes inside the LTX2.3 audio lip-sync preset:

New torchaudio-free Mel-Band RoFormer vocal separation nodes (SECoursesAudioTools: "SE Mel-Band RoFormer Model Loader" and "SE Mel-Band RoFormer Sampler"). They load the same checkpoint and give bit-identical vocals, so the LTX2 and LTX2.3 audio lip-sync presets keep their vocal separation even where the original Mel-Band RoFormer node pack cannot load.
GGUF models work again (our ComfyUI-GGUF fork): on the latest ComfyUI every quantized GGUF model failed with "unexpected keyword argument 'input_act'". Fixed and verified bit-exact on 57 test cases.
FLUX 2 Mistral GGUF text encoder loads again (ComfyUI-GGUF): the GGUF Tekken tokenizer now gives ComfyUI the special token count it needs. Token IDs match the official tokenizer.
Much faster FLUX 2 GGUF on limited VRAM (ComfyUI-GGUF): GGUF text encoders are now fully unloaded when VRAM is needed, instead of leaving 6.6 GB in VRAM. FLUX 2 Low RAM sampling went from 200-309 s/it to 1.75 s/it on a 32 GB GPU.
LTX 2.3 Foley presets fixed: ComfyUI-LTXVideo is now installed from our fork, which fixes the STG error "'AttentionTensorContainer' object has no attribute 'shape'". Existing installs switch to the fork automatically on their next update.
ControlNet Aux keeps the MIT albumentations package (our comfyui_controlnet_aux fork): upstream switched to AlbumentationsX, which is AGPL-3.0, sends usage telemetry by default and conflicts with insightface (used by ReActor).

Preset Fixes and Updates

We ran every preset with real generations on the latest ComfyUI and fixed every one that failed. The FLUX 2 Low RAM presets now free VRAM between prompt encoding and sampling:

FLUX 2 - 8 Steps Fast HQ / Realism - Low RAM: two new "Free VRAM before sampling" nodes unload the models after the prompts are encoded. Speed went from 200-309 s/it to 1.75 s/it, 35-75 seconds per image on a 32 GB GPU. The Realism Low RAM preset also had the HQ preset's guidance of 7; it is now 1, like the normal Realism preset.
Krea 2 Raw Base, Krea 2 Turbo and Krea 2 Turbo Image Edit: the saved sampler values were shifted by one position (for example steps became 1 and CFG became "euler"). Fixed.
LTX2 Image To Video and LTX2 Text To Video 8 Steps: the seed was empty, so generation failed. Fixed.
LTX2.3 Extend Any Video With Audio: the input video is now resized to a multiple of 64, so the source and the half-resolution second stage always match (1280x720 is unchanged).
LTX2.3 Foley Video To Audio: the video loader read the frame limit (169) as "skip first frames". Fixed.
LTX2.5 Text To Audio: the text encoder and prompt enhancer were connected to the wrong inputs, and the enhance prompt option held a model file name. Fixed.
LTX2.5 IC LoRA Pixel Spatial Upscaler 2x (FAST Distilled and Slower HQ): the VAE paths now point to the LTX-2 folder, where our model downloader saves them.
LTX2 and LTX2.3 Audio Lip Synch Image To Video (3 presets): now use our new torchaudio-free Mel-Band RoFormer nodes for vocal separation.
AvatarForever Unified Image or Text + Audio To Long Video: now renders an exact 832x1248 portrait by default, instead of a 768x512 pixel budget that followed the input image's shape.

Installer Improvements

Smarter node updates (Windows, RunPod and Massed Compute): the updater now asks GitHub for each node's default branch and fetches only that branch. When we move a node to a different fork, as with ComfyUI-LTXVideo in this release, existing installs switch over cleanly.
Reviewed fork updates: our forks of rgthree-comfy, ComfyUI-Easy-Use, comfyui_controlnet_aux and RES4LYF are synced with their latest upstream code. We read every new upstream commit for safety first, then re-ran our presets with identical prompts: 85 of 85 passed.
comfy-kitchen minimum version raised to 0.2.37, which the new MiniMax H3 sparse attention and long lip sync need.

Zero to AI-Ready PC in One Click: Windows & Ubuntu AI Requirements Installer Full Tutorial

Furkan Gözükara — Sun, 18 Oct 2026 00:06:22 +0000

Zero to AI-Ready PC in One Click: Windows & Ubuntu AI Requirements Installer Full Tutorial

Full tutorial link > https://www.youtube.com/watch?v=ptuLL0C72Ms

The 1-Click Windows and Ubuntu AI Requirements Installers set up your PC for AI apps such as ComfyUI, SwarmUI, music, text-to-speech and text-to-3D tools: one click on Windows, one command on Ubuntu. They install the latest NVIDIA driver, CUDA 12.9 + 13.2, cuDNN 9.20, Python 3.10-3.14, the C++ build tools, Git, Node.js, .NET SDK and FFmpeg, set PATH in the right order, test everything for real (78 Windows and 66 Ubuntu tests passed on our test PCs) and show every result in a detailed HTML report.

Chapters:

00:00:00 One click sets up your PC for AI: all tests passed
00:00:17 Windows 10, 11 and Ubuntu 22, 24, 26: ready for any AI app
00:01:03 Download the latest zip and extract it in Downloads
00:01:19 Right-click the .bat and choose Run as administrator
00:01:38 The 20 steps begin: long paths, Store Python removed
00:01:50 CUDA 12.9 + 13.2 toolkits, then cuDNN for PyTorch
00:02:14 Visual Studio 2022 Build Tools and the VC++ runtime
00:02:43 Python 3.10 to 3.14, with 3.12 as the default
00:03:12 Git, Node.js 24, .NET SDK and FFmpeg
00:03:48 PATH order, Git long paths and the page file
00:04:09 Triton cache, NVIDIA driver and the final tests
00:04:32 The HTML report: 78 tests, every step, PATH and backups
00:05:13 Transparent source files, logs.txt, then restart Windows
00:06:05 Ubuntu 22.04, 24.04, 26.04: one sudo command
00:07:00 Rerun any time to update; GCC, Clang, CMake, Ninja
00:07:46 Ubuntu report: OK messages, warnings and logs.txt
00:08:54 One-time setup, ready for our apps and other AI apps
00:09:30 Latest installer version, the guide post and support

What it installs on Windows: the latest compatible NVIDIA WHQL driver for GeForce RTX 20-series and newer (checked on every run, never downgraded); Python 3.10 to 3.14 for all users, 3.12 as default; CUDA 12.9 and 13.2 side by side (13.2 default) and cuDNN 9.20, the only cuDNN registered, so PyTorch never mixes two versions; Visual Studio 2022 Build Tools (MSVC, Windows SDK) that keep compiling after Visual Studio updates, the latest Visual C++ runtime and CUDA integration; Git, Node.js 24, .NET 10 SDK for SwarmUI, FFmpeg 9 shared and static; PATH in the right order with your own entries kept, long paths, a 100 GB page file. Ready for torch.compile, Triton and flash-attention, sage-attention, xformers and insightface builds.

On Ubuntu the C++ tools are GCC, Clang, CMake and Ninja, plus Git LFS, uv, free-threaded Python 3.13t and 3.14t and a swap file. With an NVIDIA GPU it installs the newest driver Ubuntu offers and never downgrades it. Ubuntu's own python3 stays as it is.

Windows: put the zip in Downloads (not directly on C:), extract it, right-click Windows_AI_Requirements_Installer.bat and choose Run as administrator. Nothing has to be installed first. Ubuntu: open a terminal in the folder and run sudo bash ./Ubuntu_AI_Requirements_Installer.sh (24.04 recommended).

Afterwards the HTML report opens and logs.txt is saved next to the installer. Restart your PC and open a new terminal for the new PATH. You need this setup only once, not for every app. When the post shows a newer version, run the new one: finished parts are only checked, failed parts are repaired and components are updated, with backups first (--restore-env puts them back).

Tested for real: the tests compile and start C++ and CUDA programs, build Python and PyTorch-style CUDA extensions, compile GPU code for RTX 50 cards, encode a video and build a .NET app, in the environment a new terminal gets; the optional PyTorch GPU test runs with --torch-test. Supported: Windows 10 (21H2+) and 11; Ubuntu 22.04, 24.04 and 26.04 (desktop, server, cloud GPU, WSL2, Linux Mint, Zorin OS), 64-bit x86. Without an NVIDIA GPU everything still installs and compiling works; CUDA apps need an NVIDIA GPU.

This video is for anyone who wants to run AI apps locally without installing CUDA, Python and C++ tools by hand.

For support, questions or installation issues, check the links above, the pinned comment, Patreon, Discord and the video comments, and send your logs.txt if a step fails. Thank you for watching.

1-Click Windows and Linux (Ubuntu 22, 24, 26) AI Requirements Installers - Setup Your PC and Make It Ready for All Local AI Apps, Scripts, and More

https://www.patreon.com/SECourses/posts/1-click-windows-172290854

Our AI Requirements Installers set up a Windows or Ubuntu PC for all our AI apps: one click on Windows, one command on Ubuntu. They install the latest NVIDIA driver, Python 3.10 to 3.14, CUDA 12.9 and 13.2, cuDNN 9.20, the C++ compilers, Git, Node.js, .NET and FFmpeg, set every path and variable, test everything for real and show the result in a clear report. Below you will see the real reports of a Windows 11 PC with an RTX 3090 Ti and a fresh Ubuntu 24.04 PC.

This installer will get updated as needed like when we move to Torch 2.14 and CUDA 13.4.
I have been constantly getting asked how to make initial setup easy and here you have it.
Once you run this installers, your PC will be fully ready to run all of our AI apps with perfect speed, accuracy and quality.
For Windows, recommend version is Windows 11 but should work both Windows 10 and 11
For Ubuntu, recommended version is Ubuntu 24 but should work both Ubuntu 22 and 26 too
You can run multiple times with no issues if first install fails for any reason, don't forget to restart system after running installation
This is only app that you need to run as administrator do not forget
- All of my other apps, scripts, installers are run with double click or select and hit enter
Older manual installation : https://youtu.be/DrhUHnYfwC0

Why Our AI Requirements Installers

One click on Windows, one command on Ubuntu: right-click the .bat and choose Run as administrator, or run one sudo command in a terminal. Nothing has to be installed first.
Made for our apps and their full features: ComfyUI, SwarmUI, SECourses Musubi Trainer, AI Toolkit, Ultimate TTS, Premium Upscaler Pro, Video Captioner Pro and more, with torch.compile, Triton, flash-attention, sage-attention, xformers and insightface builds.
The latest NVIDIA driver: Windows gets the newest compatible WHQL driver for GeForce RTX 20-series and newer cards, checked with NVIDIA on every run. Ubuntu gets the newest driver Ubuntu offers, R595 or newer. The driver step runs last, and a working driver is never downgraded.
Several versions side by side: Python 3.10, 3.11, 3.12, 3.13 and 3.14 with 3.12 as the default, and CUDA 12.9 and 13.2 with 13.2 as the default. Ubuntu also gets the free-threaded Python 3.13t and 3.14t.
Tested for real: 78 tests on Windows and 66 on Ubuntu compile and start C++ and CUDA programs, build Python and CUDA extensions, encode a video and build a .NET app, in the same environment that a new terminal gets. On the RTX 3090 Ti, the CUDA test programs ran on the GPU with correct results.
Fast: the Windows 11 PC with the RTX 3090 Ti was installed, updated and tested in 20 minutes 45 seconds, with the CUDA and cuDNN installers already downloaded by an earlier run.
Made for PCs that already have tools: an existing Visual Studio is updated, a current Visual C++ runtime is left alone, and an older Python from the Python Install Manager is backed up while its virtual environments keep working.
A clear HTML report: every component, step, path, variable, test, download and backup on one page, opened in your browser at the end.
The right cuDNN every time: only cuDNN 9.20 for CUDA 13 is registered, so PyTorch always loads one matching cuDNN.
Ready after updates and in every terminal: compiling keeps working after Visual Studio updates on Windows, and on Ubuntu every new terminal is ready, bash, zsh and login shells included.
Safe on your PC: every change is backed up first and one command restores it. Ubuntu's own Python stays as it is.
Run it again any time: finished parts are only checked, failed or interrupted parts are repaired, and components and the NVIDIA driver are updated.

Here are both reports right after the installation:

1 and 3: everything installed and verified, every test passed. 2: the full installation time on the Windows 11 PC with the RTX 3090 Ti. Its CUDA and cuDNN installers were already downloaded by an earlier run.

What You Download

You get small script files. They download everything else and check each download.

1: on Windows you start the .bat. It comes in Windows_AI_Requirements_Installer.zip together with Windows_AI_Requirements_Installer.py, which must stay in the same folder. 2: on Ubuntu you run the .sh. The two text files explain every step and option.

Windows AI Requirements Installer

How to Run It

Extract the zip, right-click Windows_AI_Requirements_Installer.bat and choose Run as administrator. A normal double-click works too: it asks Windows for administrator rights by itself. The .bat brings its own temporary portable Python from python.org, so nothing has to be installed first.

When it finishes, the HTML report opens in your browser and logs.txt is saved next to the .bat. Open a new Command Prompt or PowerShell window to use the new PATH. Windows is never restarted for you: restart it when the report asks, and if a new NVIDIA driver asked for the restart, run the installer again afterwards to finish the tests.

What It Installs

The report starts with every installed component, its version and its folder:

1: the latest NVIDIA WHQL driver, 617.42 on this PC. 2: Python 3.10 to 3.14 for all users in C:\Python310 to C:\Python314, and python, pip and py start Python 3.12. 3: CUDA 12.9 and 13.2 side by side with 13.2 as the default, and cuDNN 9.20. 4: FFmpeg 9 in a shared and a static build. 5: a 100 GB page file. You also get Visual Studio 2022 Build Tools with the MSVC C++ compilers and the Windows SDK, the latest Visual C++ runtime, Git, Node.js 24, the .NET 10 SDK for SwarmUI and Windows long paths.

20 Steps, Every One Timed

The Steps table shows each step with its status and time. Open the messages of any step to see exactly what it did:

On this Windows 11 PC with an RTX 3090 Ti, CUDA 12.9 and 13.2 took 5 minutes, updating its Visual Studio Build Tools 3 minutes and the five Python versions 5 minutes. PATH and the variables were backed up before they changed, the NVIDIA driver step ran last, and the tests took 1 minute 48 seconds. The CUDA installers leave out their bundled older display driver, because the driver has its own step.

The Latest NVIDIA Driver

The installer keeps your NVIDIA driver up to date. It runs as the last installation step, right before the tests:

1: the driver comes after CUDA, cuDNN and every other part, so a driver restart never comes before the other installers. 2: it finds GeForce RTX 20-series and newer desktop and laptop cards by their hardware IDs, even with no driver installed yet. 3: it asks NVIDIA's live service for the newest compatible WHQL driver on every run, downloads the official NVIDIA-signed package and installs only the display driver. 4: a driver that is already the newest is kept, as here. Use --skip nvidia to keep your driver, or --only nvidia to update just the driver.

Made for PCs That Already Have Tools

This PC already had Visual Studio Build Tools, a .NET SDK and a Python 3.14 from the Python Install Manager. The installer worked with what was there:

1: the existing Visual Studio Build Tools were updated to the latest 17.x. 2: the Visual C++ runtime was already current, so it was left alone. 3: the old Python 3.14 was unregistered with a backup, and py -3.14 now starts the new Python for all users. 4: a link keeps every existing virtual environment of the old Python working. The new .NET SDK was added next to the one already there.

PATH in the Right Order

The PATH order decides which Python, CUDA and cuDNN every app finds. The installer puts each folder at the top of the system and user PATH in the right order and shows the result in color:

Green entries were added, blue ones moved to the top, and white ones are your own entries, kept as they were. cuDNN and CUDA 13.2 come first, then the Windows system folders, then Python 3.12 and the other versions, the C++ build paths and the FFmpeg shared build. 7: the entries of your other programs stay after ours. Your PATH before the change is listed below the table.

Environment Variables

The installer sets the variables that compilers and PyTorch extension builds look for, and it shows each one before and after:

1: CUDA_HOME points to CUDA 13.2. 2: CUDNN_PATH points to the cuDNN 9.20 root that torch extension builds and CMake expect. 3: INCLUDE, LIB and LIBPATH make the C++ compiler work in every new terminal, like the Visual Studio developer prompt. 4: nvcc gets the flag it needs with the MSVC compiler. VSLANG keeps compiler messages in English for torch.compile, and values that were already correct stay unchanged.

Visual Studio Updates Keep Compiling Working

Visual Studio installs every compiler update into a new version folder. Our installer gives PATH, INCLUDE and LIB fixed paths that always lead to the current compiler and Windows SDK:

1: a fixed path leads to today's MSVC version. 2: a scheduled task points the paths at the new version at startup, at logon, after Visual Studio updates and every day, so open terminals keep working. 3: the 100 GB page file goes on the fastest suitable drive, chosen here from three drives.

78 Tests, Run for Real

At the end, the installer tests everything with exactly the PATH and variables that a new terminal or app gets:

All 78 tests passed. The first group checks that every command finds the right program: nvcc finds CUDA 13.2, cuDNN for CUDA 13 comes first, cl is the MSVC compiler, ffmpeg is the shared build, and the Windows system folders stay in front.

Every Python version is started and checked:

Each version runs with ssl, sqlite3, ctypes and tkinter, creates a virtual environment and runs pip in it, and has the headers and libraries that compiling extensions needs. The nvitop GPU monitor is installed too.

CUDA 12.9 and CUDA 13.2 are tested the way PyTorch uses them:

The tests compile, link and start a CUDA program, and on the RTX 3090 Ti its GPU result was correct with both CUDA versions. They compile the Thrust and CUB code that PyTorch extensions use, compile GPU code for RTX 50xx (Blackwell) cards and compile a kernel at run time with NVRTC. cuDNN is built into a C++ program and loads the way onnxruntime and insightface load it.

The C++ tests build what torch.compile, Triton and pip build on your PC:

An OpenMP library is built and loaded from Python, like torch.compile does on the CPU. A Python C extension is built the way pip builds insightface. CUDA 12.9 and CUDA 13.2 extensions are built with the compiler that PyTorch picks, the same way flash-attention and xformers are built. CMake and Ninja projects build too, and their CUDA programs ran on the GPU with correct results.

The last groups test FFmpeg, Git, Node.js, .NET and Windows itself:

FFmpeg encodes a test video and reads it back, and its DLLs load the way torchcodec and torchaudio look for them. A .NET console app is created, built and run, as SwarmUI needs. A file path longer than 260 characters works, and the 100 GB page file is set.

Checked Downloads and Backups

Every download is checked before it is used, and every change is backed up first:

1: every installer must carry a valid digital signature from NVIDIA, Microsoft, the Python Software Foundation or its other publisher. 2: archives pass an integrity check. 3: send us logs.txt when you ask for help. 4: the backup folder holds HOW_TO_RESTORE.txt. 5: one command puts PATH and every variable back.

Windows Options

--dry-run: shows the plan and changes nothing.
--only nvidia: checks for and installs just the latest compatible NVIDIA driver.
--skip nvidia: keeps your current driver during a full installation.
--only env,verify: runs only the steps you name. --skip skips steps, and --list-steps lists them.
--torch-test: installs PyTorch 2.13 with CUDA 13 in a temporary environment and runs real cuDNN convolution and BLAS tests on your NVIDIA GPU.
--restore-env "backup folder": puts PATH and the variables back from a backup.
--no-browser: keeps the report from opening automatically.

Pass them to the .bat from a terminal, for example:

Windows_AI_Requirements_Installer.bat --dry-run

Ubuntu AI Requirements Installer

How to Run It

It works on Ubuntu 22.04, 24.04 and 26.04 LTS: desktop, server, cloud GPU machines and WSL2, and on systems based on them such as Linux Mint and Zorin OS (64-bit x86). Put Ubuntu_AI_Requirements_Installer.sh into a folder, open a terminal there and run:

sudo bash ./Ubuntu_AI_Requirements_Installer.sh

Enter your password, and the rest is automatic. At the end, the HTML report opens in your browser, and the report and logs.txt are saved next to the .sh file. Open a new terminal afterwards, and restart Ubuntu when the report asks for it after a new NVIDIA driver. The installation keeps running even if you close the terminal.

What It Installs

The Ubuntu report starts with the same overview:

1: Python 3.10 to 3.14 and the free-threaded 3.13t and 3.14t. python and py start Python 3.12, and Ubuntu's own python3 stays as it is. 2: CUDA 12.9 and 13.2 with cuDNN 9.20 for CUDA 13 (CUDA 13.4 on Ubuntu 26.04). 3: FFmpeg 9 in a static and a shared build. 4: a swap file for 100 GB of swap in total. GCC, Clang, CMake, Ninja, Git with Git LFS, Node.js 24, the .NET 10 SDK and uv come too.

17 Steps, the NVIDIA Driver Last

Each step is listed with its time:

The installer first repairs the package system and adds the software sources for CUDA, Python and Node.js. On our fresh PC, the seven Python versions took 43 seconds and the tests 28 seconds. You also get aria2, 7-Zip, unrar, nvtop, htop and cloudflared for the apps' share links.

On a PC with an NVIDIA GPU, the driver step installs the newest driver that Ubuntu offers for the card, R595 or newer (R615 on Ubuntu 26.04), with open kernel modules from Ubuntu's signed packages. It upgrades the driver again on every run, never downgrades it, and gives RTX 50 (Blackwell) cards the open kernel module they need. It runs last, right before the tests.

Ready in Every New Terminal

The environment is written once for all users and loaded by every kind of terminal:

1: login shells, bash and zsh all load it. 2: use-cuda 12.9 switches one terminal to CUDA 12.9 when you need it. 3: CUDA_HOME, CUDA_PATH, CUDNN_PATH and the CUDA tools in PATH are set for all users. 4: this file is what every new terminal loads. The libraries are registered system wide, so no LD_LIBRARY_PATH is needed, and lines in ~/.bashrc and similar files that switch to another CUDA version are removed with a backup.

66 Tests, Run as You

Every test runs as your user with exactly the environment that a new terminal gets:

All 66 tests passed. Each Python version runs with ssl, sqlite3, lzma, readline and tkinter, creates a virtual environment with pip and compiles a C extension.

The CUDA tests build and start programs with both CUDA versions:

They compile the Thrust and CUB code of PyTorch extensions and GPU code for RTX 50xx (Blackwell) cards, check that use-cuda 12.9 switches a terminal and back, and check that every program finds the CUDA and cuDNN libraries by name, like ONNX Runtime does.

Then come the C++ and FFmpeg tests:

GCC and Clang compile C++17, an OpenMP library is built and loaded like torch.compile does on the CPU, and a GPU launcher is built the way Triton compiles one at run time. FFmpeg encodes a test video and offers the x264, x265, SVT-AV1, VP9 and NVENC encoders. Its shared libraries are found the way TorchCodec looks for them, and the files for building PyAV and decord are in place.

The last groups check the tools and your own terminal:

Git LFS, Node.js, .NET, aria2c, 7-Zip, unrar, jq, cloudflared, nvtop and htop all work, and a .NET app builds as SwarmUI needs. Your own terminal uses CUDA 13.2, and your ~/.bashrc, conda or pyenv settings keep the installed tools in reach.

Before and After, Downloads and Backups

The end of the report shows what changed and where everything is:

1: a fresh Ubuntu before, everything installed after. 2: every download is checked, the archives with SHA-256. 3: every APT package of the run is listed. 4: every changed file is backed up with HOW_TO_RESTORE.txt. 5: one command restores the files, and another removes what the installer added.

Ubuntu Options

--yes: starts without a question, for scripts and cloud machines.
--dry-run: shows the plan without changing packages or settings.
--skip and --only: skip steps or run only some, for example --skip nvidia keeps your driver (--list-steps lists them).
--swap-gb 32: sets the size of the swap file (0 = no swap file).
--torch-test: tests PyTorch 2.13 with CUDA 13 in a temporary environment: cuDNN convolution, BLAS, loaded libraries and torch.compile.
--repair: after a power loss or crash, checks every installed package and reinstalls damaged ones.
--restore-env "backup folder": puts back the files it changed.
--uninstall: removes what the installer added outside of APT.

Example:

sudo bash ./Ubuntu_AI_Requirements_Installer.sh --yes --skip swap

Run It Again Any Time

Both installers are made to run again and again. Finished parts are only checked, failed or interrupted parts are retried and repaired, and components are updated to the versions in the installer. The NVIDIA driver is checked again on every run, too. Run the newest installer whenever you download it, and your PC stays ready for our newest apps.

Better Than ElevenLabs? Ultimate TTS Voice Training & Cloning in 600+ Languages

Furkan Gözükara — Tue, 06 Oct 2026 20:27:47 +0000

Train your own AI voice with Ultimate TTS. This complete Windows tutorial covers voice cloning, automatic take selection, multilingual speech and OmniVoice full fine-tuning. The app also supports RunPod, SimplePod, Massed Compute and local Linux.

Video Chapters:

00:00:00 My trained AI voice
00:00:38 600+ supported languages
00:01:19 Windows and cloud options
00:01:46 Your first voice clone
00:02:10 Generate and listen
00:03:18 Download ZIP and guides
00:03:46 Windows Requirements
00:04:28 Choose an install path
00:04:40 Extract the package
00:04:54 Run the installer
00:05:23 PyTorch, CUDA and wheels
00:05:54 Start the application
00:06:19 Models, voices, GPU presets
00:07:04 First load vs warm runs
00:07:46 Players and saved outputs
00:08:34 IndexTTS English example
00:09:02 AuK English example
00:09:52 Reference library and files
00:10:48 Select a reference range
00:11:38 Microphone recording
00:11:53 Reference text and transcription
00:13:19 Clone and Auto voice
00:14:38 Voice design tags
00:15:26 Language selection
00:15:40 Preview and Smart sentences
00:16:32 Section-length study
00:17:01 Sentence and section pauses
00:18:08 Pause tags
00:19:04 Check and set pronunciation
00:19:38 Save and reuse dictionary
00:21:10 Expressive tags
00:21:44 Keep which take
00:21:58 Fewest word errors
00:22:12 Rank takes and set checks
00:22:42 Ranking and fallback result
00:23:49 Whole-text candidates
00:24:36 60-line selection study
00:25:05 139-line Auto voice study
00:25:35 First vs selected take
00:26:42 Matching-language references
00:27:26 Arabic voice example
00:28:24 Chinese voice example
00:29:04 French voice example
00:29:46 German voice example
00:30:40 Italian voice example
00:31:22 Japanese voice example
00:32:03 Korean voice example
00:32:58 Polish voice example
00:33:38 Spanish voice example
00:33:49 Cross-language cloning
00:35:48 Speaking-rate comparison
00:36:49 Target duration modes
00:37:39 Best quality: 32 steps
00:38:25 Fast: 16 steps
00:38:54 Seeds and advanced sampling
00:39:44 WAV, MP3 and filenames
00:40:01 Reference and subtitle export
00:40:16 Inspect the exported files
00:41:10 Recent outputs as references
00:41:52 Audio tuning comparison
00:43:09 Tuning and edge silence
00:43:30 SRT import and cue timing
00:44:22 Create a narrated MP4
00:45:08 Batch inputs and filenames
00:45:43 Shared and per-file voices
00:46:07 Batch execution and progress
00:46:47 Inspect batch results
00:48:00 Save a personal preset
00:48:18 Load settings and reference
00:49:26 Recent settings and theme
00:49:45 Select GPU and memory tier
00:50:02 BF16 and ConvRot INT8
00:50:16 Measured memory use
00:50:30 Measured synthesis time
00:50:44 Compare both variants
00:52:25 Section batch size
00:52:39 Load models and find files
00:53:03 VRAM benchmark and tier cap
00:53:56 Train a new OmniVoice
00:54:15 Source recordings and text
00:54:44 Reserve evaluation sessions
00:55:00 Dataset folder and scan
00:55:37 Transcript and alignment
00:55:57 Segment duration
00:56:13 Padding and caption cleanup
00:56:28 Quality filters
00:56:47 Prepare and inspect clips
00:58:02 Speaker and transcript audit
00:58:50 Review flagged clips
00:59:21 Correct words and boundaries
01:00:20 Build the OmniVoice cache
01:00:47 Dataset and training name
01:01:06 Full fine-tuning
01:01:33 Full, LoRA and DoRA
01:01:50 Held-out voice comparison
01:02:05 Epochs and Training plan
01:02:22 Token budget and accumulation
01:02:39 Learning rate and memory
01:03:14 Checkpoints and train state
01:03:30 Continue vs weights-only
01:03:47 Evaluation and final test
01:04:18 Fluency analysis and filters
01:05:16 Start training
01:05:49 Speed, loss and validation
01:06:18 Hear a training sample
01:06:45 Stop and resume controls
01:07:05 800-update completion
01:07:39 Checkpoint analysis
01:07:55 Automatic speech comparison
01:08:15 Reference audition
01:08:37 Speaking-rate calibration
01:08:53 Save an INT8 voice copy
01:09:46 Load the trained preset
01:10:14 Generate held-out speech
01:11:12 Base vs trained voice
01:12:15 Configure Checkpoint Grid
01:13:24 Hear the six grid cells
01:14:34 Trained Auto voice
01:15:22 Trained INT8 voice
01:16:13 Updates, backup and help
01:17:01 Your next voice

This video is for anyone who wants a complete voice workflow, from a short reference to a reusable trained voice. Start with the included reference and save your first working preset.

For support, questions, feature requests, installation issues and updates, use the download post, pinned comment, Patreon, Discord and video comments. Thank you for watching.

Ultimate Text To Speech Generator With Voice Cloning turns your scripts, subtitles and whole chapters into natural speech on your own PC, with three speech models in one app: IndexTTS 2.5, OmniVoice and AuK. Clone any voice from a short clip, design a new voice from a few words, edit recordings, and train your own voice from your videos. Below you will see every feature with real screenshots from the app.

OmniVoice DoRA (better version of LoRA) trainings is yielding literally better than ElevenLabs with our special pipeline + Keep which take inference system.

Why Ultimate Text To Speech Generator

Three speech models in one app: IndexTTS 2.5, OmniVoice and AuK. Switch them in the header; each keeps its own settings.
Clone any voice in seconds: drop an audio or video file and type your text. On an RTX 5090, IndexTTS 2.5 made 31 seconds of speech in 18 seconds from a 14.6 second part of a video.
Take quality: the app renders several takes of every section, Whisper checks them, and the take that sounds most like the voice without a word error is kept. In the developer's test on 60 lines, word errors fell from 0.65 % to 0.20 % with OmniVoice.
New voices without a recording: describe a voice in one sentence with AuK, or pick voice tags with OmniVoice.
AuK Audio Editing: 23 editing tasks: replace, add or remove words, change speed, pitch, volume or emotion, clean up noise and separate voices. Replacing two words in a 10 second clip took 1.7 seconds.
Train your own voice for any model: from your videos and subtitles to a finished voice with a ready-to-use preset. Four recordings became 205 clean training clips in 50 seconds.
Made for long narration: Smart sentences, exact pause tags, a live section preview and a pronunciation dictionary.
Subtitles to speech and MP4: 10 subtitle formats, speech timed to every cue, and an MP4 with your image in the same run.
Batch generation: whole folders of scripts and subtitles in one click.
Ready for your GPU: 7 GPU presets from 6 GB to 32 GB, BF16 or INT8 ConvRot models, block swap and a built-in VRAM benchmark.
1-click installers: Windows, RunPod, SimplePod, Massed Compute and Linux, with PyTorch 2.14.1 and CUDA 13. The models download automatically, and every model weight file is checked with SHA256.
Frequent updates: four releases between 2 and 4 October 2026, with a built-in Help tab and changelog.

Here is the app right after a voice clone. The voice came from a 14.6 second part of one of our tutorial videos, and 31 seconds of speech were ready in 18 seconds.

1: choose IndexTTS 2.5, OmniVoice or AuK in the header. 2: the preset for your GPU is selected on the first start. 3: any audio or video file can be the voice. 4: type your script. 5: the result plays right away, with the elapsed time, speed and VRAM next to it.

1-Click Installation

What You Download

You get one small zip file with the installers for every platform.

Extract it into a short folder path without spaces, double-click Windows_Install_or_Update.bat, then start the app with Windows_Start_App.bat. Run the same installer again at any time to update.

Latest PyTorch and CUDA 13

The installer makes its own Python 3.12 virtual environment, so your other apps stay untouched. It installs PyTorch 2.14.1 with CUDA 13, Transformers, Triton for Windows and every other package with uv, a very fast package installer.

PyTorch, torchvision and torchaudio always come from the same CUDA 13 build. Our fresh install took 5 minutes 39 seconds on our PC, models included.

Automatic Model Downloads

At the end, the installer downloads the IndexTTS 2.5 models and their helper models by itself.

Large files download over 16 connections, and every model weight file is checked with SHA256, so a broken download is caught right away. Running the installer again skips the large files you already have.

Start the App

Double-click Windows_Start_App.bat. On the first start, the app reads your GPU memory and selects the matching GPU preset.

The app was ready in 9.5 seconds on the first start and in 4 to 6 seconds later. OmniVoice, AuK and the built-in Whisper download automatically the first time you use them.

Cloud GPUs: RunPod, SimplePod and Massed Compute

You can also run the app on a cloud GPU. The cloud installers set up everything with one command and download the models for you. A Gradio share link lets you use the app from any device.

SimplePod: register here and use this template.
RunPod: register here and use this template.
Massed Compute: register here and use our coupon SECourses.

The step-by-step commands are in the instruction files inside the zip.

Requirements

On Windows you need Python 3.12, Git, FFmpeg, CUDA 13, cuDNN 9.17 and Visual Studio with the C++ tools. This tutorial shows every step, and the same setup runs all our AI apps.

Three Speech Models in One App

Every model has its own strengths, and you switch between them with one click in the header.

IndexTTS 2.5 gives you emotion control and reads five languages. OmniVoice clones voices, designs voices from tags and speaks many languages. AuK clones voices, designs voices from a plain sentence and edits recordings in its own tab. Each model keeps its own settings, and every preset stores all three.

GPU Presets for 6 GB to 32 GB Cards

Seven read-only presets fit the whole app to your card: generation, dataset preparation and training together. On the first start, the app selects the preset of your GPU.

Where to find it: the Universal preset list sits at the top of the page, above every tab.

Open the Universal preset list to see every preset:

Change anything you like, type a name and click Save to keep your own preset. It stores every setting of every tab and of all three models. Every training adds a ready-to-use preset of its new voice to this list. The Help tab shows what each GPU preset uses:

A smaller card pays with speed first: helper models move to the CPU, and the 6 GB preset streams 22 of the 24 GPT blocks from RAM.

Voice Cloning From Any Audio or Video

A clean clip of a few seconds is enough to clone a voice. Audio and video files both work, and the preview plays exactly what the app will use.

Where to find it: open the Voice Generation tab. Reference Voice is the left column.

Here is the Reference Voice column up close:

Drop a file, record from your microphone, or load any file on your PC by its path. Time ranges such as 0.2:14.8 keep only the clean part of a long recording, and several ranges are joined in order.

Text, Smart Sentences and Pauses

Long text is split into sections before the voice reads it. Smart sentences packs whole sentences into each section.

Where to find it: in the Voice Generation tab, Text & Timing is the middle column.

Here are the text controls:

Tags such as [pause:800ms] insert an exact silence. With a trained voice, the line length and the pauses between sentences come from the speaker's own recordings. Below the text, the live section preview shows every section before you generate:

You see the tokens and words of each section and every pause, so you can fix the text before you spend any GPU time.

Pronunciation Check and Dictionary

Names and technical words are the hardest part of any narration. The pronunciation check finds the words the model may not know and suggests a reading for each.

Where to find it: in the Voice Generation tab, click Pronunciation check & dictionary under the live section preview.

Click Check unknown words:

Readings come from the CMU dictionary, acronym and CamelCase splitting, and letter rules. Save the good ones to your dictionary, and the app applies them in Voice Generation, Batch Generation and the live preview, for IndexTTS 2.5 and OmniVoice.

Take Quality: The Best Take of Every Section

Take quality picks the best take of every section. It renders several takes, lets the built-in Whisper check every word, and keeps the take that sounds most like the voice without a word error. It works for all three models, and the GPU presets switch it on.

Where to find it: in the Voice Generation tab, Take quality is in the right column, under the live log.

Here is Take quality after an OmniVoice clone of the same video clip:

OmniVoice rendered 10 takes of our 32 second script in 26.5 seconds, and the app kept the most similar take with 0 word errors. In the developer's test on 60 lines of voice cloning, word errors fell from 0.65 % to 0.20 % with OmniVoice and from 0.91 % to 0.35 % with IndexTTS 2.5.

OmniVoice: Clone, Design or Auto Voice

OmniVoice clones a reference voice, designs a new voice from tags, or picks a voice by itself.

Where to find it: choose OmniVoice in the header. Its section opens in the Voice Generation tab, under Voice LoRA / DoRA.

Here is the OmniVoice section with a designed voice:

We picked four tags: female, young adult, moderate pitch and british accent. The new voice spoke 6.1 seconds of speech in 1.2 seconds, with no reference recording at all.

AuK: Describe a Voice in Words

AuK clones voices, designs a voice from one plain sentence, and lets a trained AuK voice speak without a reference.

Where to find it: choose AuK in the header. Its section opens in the Voice Generation tab, under Voice LoRA / DoRA.

Here is the AuK section with a described voice:

"A calm middle-aged man with a deep, warm voice, speaking slowly and clearly, like a documentary narrator." AuK made 6.6 seconds of speech from that sentence in 1.5 seconds. AuK speaks English and Chinese.

Emotion Control

IndexTTS 2.5 separates the voice from the emotion, and the app gives you four ways to set the emotion.

Where to find it: with IndexTTS 2.5 selected, scroll down in the Voice Generation tab and click Emotion Control. The Open / close all sections button at the top opens it too.

Here is Emotion Control with the emotion sliders selected:

Keep the speaker's own tone, copy the delivery of an emotion clip, mix joy, anger, sadness, fear, disgust, depression, surprise and calm, or describe the emotion in words. Emotion weight sets how strong it is.

Timing, Output Formats and Word Timestamps

Make the speech fit your video, save it in the format you need, and get subtitles of every word.

Where to find it: in the Voice Generation tab, click Segmentation & Timing, Output and Execution.

The three sections opened:

Ask for an exact length, such as 25 seconds, and Natural mode paces the speech to fit. Every run saves a WAV file, one tick adds an MP3 up to 320k, and another adds SRT and VTT subtitles with the timing of every word. Isolated subprocess mode frees all VRAM after each job.

Subtitles to Timed Speech and MP4

Turn any subtitle file into speech that follows its timing, ready for dubbing, translations and slide videos.

Where to find it: in the Voice Generation tab, the Captions and Still image for MP4 boxes are under the text controls.

Drop your subtitle file and tick Use caption cue timing:

SRT, VTT, SBV, ASS/SSA, SUB, LRC, TTML/DFXP, SAMI, JSON and TSV all work, and the app recognizes the format from the file itself. Add an image, and the same run also makes an MP4:

Our 6-cue subtitle file became exactly 30.00 seconds of speech, with every cue starting on time.

AuK Audio Editing

AuK edits real recordings: change the words, the delivery or the sound, and separate voices.

Where to find it: choose AuK in the header, then click the AuK Audio Editing tab.

Here is a word replacement:

We replaced "quick test" with "short demo" in a 9.9 second recording. The edit took 1.7 seconds, and Whisper heard the new words in the result. The 23 tasks also insert or remove words, change speed, pitch, volume and emotion, turn speech into a whisper, add a laugh, clean up noise and reverberation, and keep one speaker or the voices from music.

Batch Generation

Generate many scripts or subtitle files in one run.

Where to find it: click the Batch Generation tab.

Drop your files, paste your texts or point the app at a folder, then click Generate batch:

Our three stories became 116 seconds of speech in 65 seconds with OmniVoice, every section picked from 10 takes. Use one voice for all files, or a matching audio file next to each script for a different voice per file. Continue after item errors keeps the batch going.

Train Your Own Voice

A trained voice follows the speaker much more closely than a single reference clip: the pace, the pauses and the way each word is said. You can train IndexTTS 2.5, OmniVoice and AuK, and the app takes you from videos and subtitles to a finished voice.

Prepare the Dataset

Point the app at your videos or audio files with their subtitles. It cuts them into clean clips of whole sentences.

Where to find it: click the LoRA Dataset Preparation tab.

Here is a real run on four of our tutorial recordings:

Your subtitles give the words and Whisper gives the exact word timing. The four recordings, 46.8 minutes in total, became 205 clean clips of 40.6 minutes in 50 seconds. Every clip can be checked before training:

The clips are 4 to 16 seconds long, and every clip keeps its text.

Voice and Transcript Audit

The audit checks every clip against a clean recording of the speaker and against its text, and builds a separate training dataset from the clips that pass.

Where to find it: in the LoRA Dataset Preparation tab, click Voice and transcript audit at the bottom.

Here is the audit of our 205 clips:

It checked 205 clips in 1 minute 14 seconds and kept 194. A second Whisper model rescued 7 good clips that the first one misheard. A whole recording can be held out, so the app tests the voice on speech it never trained on.

Start the Training

Choose your dataset, give the voice a name and click Start training. The defaults are the settings we measured on real voices.

Where to find it: click the Voice Training tab.

Here is the tab with our audited dataset:

The GPU VRAM preset fits the training to your card. DoRA and LoRA train small adapters; full fine-tuning trains the speech model's own weights on 16 GB cards and larger (AuK 32 GB). The fluency filter can train on fluent clips only. During training the dashboard updates live:

Our IndexTTS 2.5 voice trained at about 5 to 6 steps per second with 2.3 GB of VRAM. The app checks the voice on held-out clips every epoch and keeps the best checkpoint.

Automatic Checks After Training

When training ends, the app keeps working for you. It renders the same sentences with the base model and with the checkpoints, then compares them with the speaker's real recordings of those sentences. It also calibrates the speaking rate, trains a voice decoder adapter, tunes the decoding settings and auditions reference clips. Here is the comparison for our voice:

Likeness to the speaker's real recordings rose from 0.810 with the base model to 0.867 for the chosen checkpoint with its voice decoder adapter, and the word error rate fell from 2.4 % to 2.3 %. The app picked that checkpoint by itself and set it as the voice's default.

A Ready-to-Use Preset

Every training ends with a ready-to-use preset of the new voice: our training saved my_voice_Takes_5 in the Universal preset list you saw above. It loads the chosen checkpoint, the best reference clip from the audition, the calibrated speaking rate of 0.86, the decoding settings from the sweep, and Take quality with up to 5 takes per section.

Use Your Trained Voice

Select your voice in Voice Generation, and the app loads everything it learned.

Where to find it: in the Voice Generation tab, the Voice LoRA / DoRA panel is right under the three columns. You can also load the voice's ready-made preset from the Universal preset list.

Here is the panel after loading the preset of our new voice:

The panel shows the voice's calibrated pace, the line length and the pauses measured from its own training clips, and the decoding its sweep adopted, and applies them to your text. With this preset our voice read the 88-word script as 35.2 seconds of speech in 28 seconds, and the built-in Whisper found 0 word errors in every section.

Compare Checkpoints by Ear

The Checkpoint Grid tab renders the same texts with every checkpoint you choose, so you can listen to them side by side.

Where to find it: click the Checkpoint Grid tab.

Here is a grid of our voice with the base model and three checkpoints:

Every cell uses the same texts, reference and seed, so you hear only the difference between the checkpoints. Every grid is saved and opens again at any time, and Calibrate speaking rate from this grid measures the voice's real pace.

Models and Performance

Runtime

Fit the models to your GPU in one place.

Where to find it: click the Models & Performance tab.

Here are the runtime controls:

Choose a VRAM tier, and the fit check shows the expected peak memory before anything loads. Pick the BF16 model or the INT8 ConvRot model, which saves memory. Unload model / free VRAM clears the GPU in one click.

Block Swap and Model Placement

Small cards can stream parts of the model from system RAM.

Where to find it: in the Models & Performance tab, click Block Swap & Memory and Auxiliary Model Residency.

Here is the 6 GB tier:

The 6 GB tier streams 22 of the 24 GPT blocks from RAM and keeps the reference models on the CPU. Every helper model can stay on the GPU, stay on the CPU or load on demand.

Model Files and VRAM Benchmark

Check your model files and test a GPU tier before a long job.

Where to find it: in the Models & Performance tab, click Model Files & Downloads and VRAM Benchmark.

Both sections after a real test:

The INT8 ConvRot GPT is 1.1 GB instead of 3.1 GB and downloads from inside the app. The VRAM benchmark ran the 8 GB tier with its memory limit in place on our card: it fit with a 3.5 GB peak and made 25.9 seconds of speech in 19.3 seconds.

Built-in Help and Changelog

You never need to leave the app to learn it.

Where to find it: click the Help tab or the Changelog tab.

The Help tab starts with a four-step quick start:

It explains every model, every workflow and every parameter. The Changelog tab lists every release in plain words:

The newest release is always at the top.

Latest Updates

The app gets frequent updates. Versions 7.0 to 8.1 came out between 2 and 4 October 2026:

8.1: fresh installs keep PyTorch, torchvision and torchaudio on one CUDA 13 build, the preset after training carries the voice's calibrated speaking rate and decoding, IndexTTS training caches its features by itself, and the chosen speech model stays selected after a reload.
8.0: Take quality for all three models, a ready-to-use preset after every training, and OmniVoice as the default model of a new install.
7.1: AuK, the third speech model, with voice design from a sentence, trained voices without a reference, and the AuK Audio Editing tab.
7.0: OmniVoice, the second speech model, with voice design, BF16 or INT8 ConvRot and training.

Get Ultimate Text To Speech Generator

Download the latest zip file attached to this post, extract it and run the installer for your platform. To update, get the newest zip, overwrite the old files and run Windows_Install_or_Update.bat again.

Requirements tutorial: Python, Git, FFmpeg, CUDA and C++ tools step by step.
Support: our Discord channel.

Whisper-WebUI Premium Transcribes a 10-Minute Video in 4 Seconds: Full Windows Guide

Furkan Gözükara — Thu, 01 Oct 2026 22:54:02 +0000

App download link > https://www.patreon.com/SECourses/posts/whisper-webui-to-145395299

Full tutorial link > https://www.youtube.com/watch?v=yzNhkC3aQuA

Whisper-WebUI Premium turns video and audio into subtitles, searchable transcripts and timed data on your own Windows PC. See a real 10-minute recording transcribed in 4 seconds, then follow the complete fresh installation and every main workflow. This tutorial covers all 3 engines, researched quality presets, automatic model downloads, 6 export formats, custom vocabulary, folder batches, YouTube sources, microphone recording, speaker labels, subtitle translation and voice/music separation.

Links:

Download Whisper-WebUI Premium and read the full feature guide: [ https://www.patreon.com/SECourses/posts/whisper-webui-to-145395299 ]

Windows Requirements tutorial: [ https://youtu.be/DrhUHnYfwC0 ]

You will learn how to download the latest ZIP, locate it in Chrome, move it to a permanent folder and extract the whole package. Read the Windows Requirements section and follow its linked setup tutorial first if needed. Use an English-character path without spaces, such as G:\whisper_tutorial. Run the Windows installer and app starter by double-clicking their BAT files or selecting them and pressing Enter; never run them as administrator. Keep the start window open while using the local browser interface. We then load a video by upload or full file path, choose a language and output formats, generate the first subtitles, and inspect the saved files and transcript. Use SRT for conventional video subtitles, WebVTT for compatible web players, and TXT for searching or rewriting the transcript. LRC, JSON and TSV keep timed information for other tools. Select several formats in the same run and inspect matching filenames in the outputs folder.

Main topics include multilingual Whisper and optimized models, Canary-Qwen for English speech, Insanely Fast Whisper, GPU-aware presets, INT8 ConvRot, PyTorch 2.13 and CUDA 13 precompiled packages. Save your own preset with vocabulary, reload its settings, reset engine defaults or delete a user preset. Inspect compute precision, beam search, prompts and hotwords, repetition and silence controls, batch size, model offloading, word timestamps and subtitle punctuation. See a long lecture result and stop a separate worker with Cancel Generation. The filter demonstrations cover background music removal, Silero voice detection and diarization for a two-man conversation. Batch Processing scans nested input folders, mirrors their structure in a separate output folder, skips existing results and offers deliberate overwrite. The YouTube tab reads a video link or processes recent videos from a channel. Live Mic previews incoming text; Record Then Generate creates subtitles after recording. Local NLLB translates an English SRT to Spanish while preserving all 10 cue timings and speaker labels; the DeepL API controls provide another route with your own key. BGM Separation saves individual voice and instrumental tracks. SRT, WebVTT, TXT, LRC, JSON and TSV can be selected together and downloaded in one ZIP. Finally, switch themes, open or close the settings panels, and follow the package update route.

Use the chapters to follow the fresh Windows install or jump straight to engines, presets, batch folders, speaker labels, microphone modes, translation and audio separation.

Chapters:

00:00:00 A ten-minute video subtitled in four seconds

00:01:09 See the transcript and choose the right exports

00:01:48 Download the ZIP and check Windows requirements

00:02:43 Move and extract the complete installer package

00:03:32 Fresh Windows installation with the BAT installer

00:04:46 Launch the app and make your first subtitles

00:07:10 Three engines and the long lecture result

00:09:20 Save and reload your own working preset

00:10:34 Useful advanced controls without rebuilding your setup

00:15:24 Music removal, voice detection and speaker labels

00:17:35 Transcribe nested folders and reuse finished outputs

00:19:10 Video links and whole-channel processing

00:20:16 Live preview and record-then-generate microphone modes

00:21:53 Translate subtitles and preserve their timings

00:23:10 Save separate voice and instrumental audio tracks

00:23:59 Updates, cloud support and the complete workflow

The opening result is a 10-minute 56-second recording completed in 4 seconds on an RTX 5090, using the Canary Qwen Best Quality preset, batch size 16 and the optimized Canary INT8 ConvRot model with the model and caches ready. The app reports the elapsed time and result on screen.

The package also supports RunPod, SimplePod and Massed Compute for cloud GPU workflows. This video teaches Windows installation and use; the included platform instructions cover those cloud routes.

Read the linked post for release notes, illustrated feature explanations and current downloads. The app link is also in the pinned comment. For questions, installation support, feature requests and feedback, use the post and video comments.

Whisper-WebUI Premium turns any video or audio into accurate subtitles and transcripts on your own PC. On an RTX 5090, our own INT8 ConvRot engine transcribed a 1 hour 29 minute lecture in 32 seconds. Every setting comes ready with researched best-quality presets. Below you will see every feature with real screenshots from the app.

Why Whisper-WebUI Premium

Our own INT8 ConvRot engines: Whisper large-v3 runs 3.4x faster and NVIDIA Canary-Qwen 2.5B runs 8.9x faster than the standard models, on the same video and the same GPU.
Same accuracy as the original models: measured through the app on 14 public English test sets, 48 hours of audio.
Real speed: a 10 minute video in 9 seconds, a 1.5 hour lecture in 32 seconds.
3 engines in one app: Whisper, Insanely Fast Whisper and NVIDIA Canary-Qwen 2.5B, with 21 Whisper models and 100 languages.
Ready presets: researched best-quality settings for every engine, plus your own saved presets.
6 output formats in one run: SRT, WebVTT, TXT, LRC, JSON and TSV, with a one-click ZIP download.
Everything in one place: batch folders, YouTube links and whole channels, live microphone, speaker labels, background music remover, voice detection filter and subtitle translation to 200 languages.
1-click installers: Windows, RunPod, SimplePod, Massed Compute and Linux, with PyTorch 2.13, CUDA 13 and precompiled Flash Attention, xFormers, SageAttention and Triton.
Automatic model downloads with live progress, frequent updates and support.

Here is the app right after a job. A 10 minute 54 second video was transcribed in 9 seconds, 70 times faster than real time.

1: pick a ready preset. 2: download all subtitle files in one ZIP. 3: preview your video instantly. 4: watch the transcription live. 5: choose from 3 engines and our INT8 ConvRot models.

Speed: Our INT8 ConvRot Engines

We built INT8 ConvRot versions of Whisper large-v3, Whisper large-v1 and NVIDIA Canary-Qwen 2.5B, and our own GPU engine to run them. The model files are smaller too: 1.6 GB instead of 3.1 GB for Whisper, and 2.9 GB instead of 5.1 GB for Canary-Qwen. Here is the same video on the same RTX 5090 with the same settings:

Long files are just as fast. Our 1 hour 29 minute lecture took 1 minute 15 seconds with our INT8 Whisper large-v3 (71x real time) and only 32 seconds with Canary-Qwen (168x real time). This is the Canary-Qwen run in CMD:

After each job the model waits in RAM and your VRAM is free. The next job moves it back to the GPU in under a second, so it starts right away.

Same Accuracy as the Original Models

Speed only matters with accuracy. We measured every model through the app on 14 public English test sets with human transcripts: 2,700 short clips and 120 long recordings, 48 hours in total.

On the Open ASR Leaderboard's 8 English test sets, our INT8 models reach the published accuracy of the full-precision originals. Canary-Qwen 2.5B scored 5.62% word error rate against the published 5.63%, and Whisper large-v3 scored 7.22% against 7.44%.

1-Click Installation

What You Download

You get one small zip file with the installers for every platform.

Extract it into any folder. On Windows, double-click Windows_Install_Update.bat, then start the app with Windows_Start_app.bat. Run the same installer again at any time to update.

Latest PyTorch, CUDA 13 and Precompiled Libraries

The installer makes its own Python 3.12 virtual environment, so your other apps stay untouched. It installs PyTorch 2.13 with CUDA 13 and our precompiled Flash Attention, xFormers, SageAttention and Triton for Windows. You never compile anything.

Every package version is tested with the app. This is the end of a real fresh install on our PC:

At the end, the installer downloads the speaker label models from our mirror. You do not need a Hugging Face token or any model approval.

Start the App

Double-click Windows_Start_app.bat. CMD shows every startup step with its time.

On our RTX 5090 the app was ready in 6.5 seconds. Open the local address in your browser and start transcribing.

Cloud GPUs: RunPod, SimplePod and Massed Compute

You can also run the app on a cloud GPU. The cloud installers set up Python 3.12, FFmpeg n9.0 and everything else with one command, and a Gradio share link lets you use the app from any device.

SimplePod: register here and use this template.
RunPod: register here and use this template.
Massed Compute: register here and use our coupon SECourses.

The step-by-step commands are in the instruction files inside the zip.

Requirements

On Windows you need Python 3.12, Git, FFmpeg, CUDA 13, cuDNN 9.17 and Visual Studio with C++ tools. This tutorial shows every step, and the same setup runs all our AI apps. The app runs on NVIDIA GPUs from the GTX 16 and RTX 20 series up to the RTX 50 series. Our INT8 ConvRot engines use the RTX 30 series and newer, and other GPUs switch to the standard models automatically.

Config Presets

Every engine comes with a locked best-quality preset. We researched each value on real test sets, so you get top results without changing a single setting.

Where to find it: Config Presets sits at the top of the page and stays there on every tab.

Open the Select Preset list to see every preset:

Change anything you like, type a name and click Save to keep your own preset. The app remembers your last preset and loads it at every start.

Three Engines and 21 Whisper Models

Choose the engine in Base Model: Whisper (faster-whisper), Insanely Fast Whisper (Transformers) or NVIDIA Canary-Qwen 2.5B. Each engine loads its best settings when you select it.

Where to find it: open the File tab and scroll to Base Model. The Model list is right under it, and the Youtube and Mic tabs have the same controls.

Open the Model list to see every model:

You get every Whisper model from tiny to large-v3, plus turbo, distil and our two INT8 ConvRot models. A model downloads automatically the first time you use it.

NVIDIA Canary-Qwen 2.5B

Canary-Qwen 2.5B is one of the most accurate open English speech models on the Open ASR Leaderboard (September 2026). Our INT8 ConvRot build runs it with Triton kernels and CUDA graphs.

Where to find it: pick Canary Qwen Best Quality in Select Preset, or choose Canary-Qwen (NVIDIA NeMo) in Base Model.

The Canary settings then load automatically:

The app reads your GPU memory at startup and picks the batch size for you: batch 2 on 6 GB, 4 on 8 GB, 8 on 10 to 12 GB and 16 on 16 GB and more. Here it chose batch 16 on the 32 GB RTX 5090 and finished the 1 hour 29 minute lecture in 36 seconds.

Automatic Model Downloads

You never download models by hand. The first time you use a model, the app downloads it and shows the progress in Live Transcription and in CMD.

All models and caches stay inside the app's own models folder, so the rest of your PC stays clean.

100 Languages and Translation to English

Whisper understands 100 languages, and Automatic Detection finds the language for you.

Where to find it: in the File tab, the Language list and the Translate to English checkbox sit next to the Model list.

Open the Language list to see every language:

Turn on Translate to English and Whisper writes English subtitles directly from speech in any language.

Output Formats and Run Controls

All the main controls sit in one clear panel under the model settings.

Where to find it: in the File tab, right under the Model row. Tick your formats, then click GENERATE SUBTITLE FILE.

Here are the formats and run controls up close:

1: tick as many formats as you like. 2 and 3: start and stop jobs. 4: previous-text context is tuned for long files automatically. 5: open the extra sections for advanced settings, the music remover, the voice filter and speaker labels. 6: every job ends with a clear summary. One run writes all six files, named after your input file:

The best-quality presets use word timestamps. You get clean sentence-level subtitles, or word-level highlighted SRT and WebVTT when you want them.

Advanced Parameters

Where to find it: open the File tab, scroll down and click Advanced Parameters. The Open / close all sections button at the top opens it too, together with every other section.

The panel opens with every decoding setting, each with a short explanation under it:

Use Hotwords or the Initial Prompt for names and terms. Turn on Use Batched Inference for extra speed on long files. Offload Models to RAM When Idle keeps your models ready while your VRAM stays free.

Music Remover, Voice Filter and Speaker Labels

Three built-in filters prepare your audio before transcription.

Where to find it: in the File tab, scroll down. The three sections sit right under Advanced Parameters; click a title to open it.

Here are the three sections opened:

The Background Music Remover separates speech from music with UVR MDX-Net. The Silero voice filter skips silence. Diarization adds speaker labels such as SPEAKER_00 and SPEAKER_01 without a Hugging Face token. In our test it labelled every turn of a two-person interview correctly.

Batch Processing Whole Folders

Transcribe a whole folder in one click, subfolders included.

Where to find it: in the File tab, on the right side, next to the upload box.

Tick Enable Batch Processing and set your folders:

Here 3 videos in 3 subfolders, 18 minutes in total, were finished in 26 seconds. The output folder mirrors your folder tree, finished files are skipped on the next run, and one broken file never stops the batch.

YouTube Videos and Whole Channels

Paste a YouTube link and the app loads the thumbnail, title and description.

Where to find it: click the Youtube tab. The Youtube Link box is at the top.

The link loads the video details right away:

Click Generate and the app downloads the audio and transcribes it. Our 12 minute video went from link to finished subtitles in about 15 seconds. Turn on Mass Transcribe Latest Channel Videos to process the latest videos of a whole channel, one after another.

Live Microphone

Speak and watch the text appear. The Mic tab has two modes.

Where to find it: click the Mic tab. Live Mic is on the left and Record Then Generate on the right.

Press Live Mic Record and start speaking:

Live Mic shows a preview while you speak and updates it every 2 seconds. When you stop, the whole recording is transcribed with the full model and saved as subtitle files. Record Then Generate lets you record first and create the subtitles after.

Subtitle Translation to 200 Languages

Translate your subtitle files with Meta NLLB, right inside the app.

Where to find it: click the T2T Translation tab, drop your subtitle files, then open the NLLB tab. DeepL API is right next to it.

Choose the languages and click TRANSLATE SUBTITLE FILE:

Choose the source and target language, click Translate and download the new file. Timestamps and speaker labels stay in place. A DeepL API tab is included too, for your own DeepL key.

Background Music Separation

The BGM Separation tab splits any audio file into music and voice.

Where to find it: click the BGM Separation tab, drop your audio files and click SEPARATE BACKGROUND MUSIC.

After the separation you get two tracks:

You get a music-only track and a clean voice track, both saved in your outputs folder.

Light and Dark Theme

The app starts in dark mode, and one click switches to the light theme.

The Open / close all sections button opens or closes every panel at once.

Latest Updates

Whisper-WebUI Premium gets frequent updates. Versions 12.4 to 13 came out between 27 and 29 September 2026:

New INT8 ConvRot models for Whisper large-v3, large-v1 and Canary-Qwen 2.5B, downloaded automatically.
Lower word error rate in English: the INT8 models now match the published accuracy of the originals.
Canary-Qwen transcribes recordings up to 40 seconds in one piece and cuts longer ones at pauses.
Canary batch size is picked from your GPU memory, with automatic recovery when VRAM runs short.
Models wait in RAM between jobs, so the next job starts in under a second.
Faster start, with every startup step shown in CMD.
Batch processing keeps your subfolders, skips finished files and lists any failed files at the end.
All GPU jobs wait in one queue and run one after another, so two jobs never clash.
A redesigned interface with light and dark themes.

Get Whisper-WebUI Premium

Full tutorial video: watch it on YouTube.
Requirements tutorial: Python, Git, FFmpeg, CUDA and C++ tools step by step.
Support: our Discord channel.

IndexTTS-2.5 SECourses Premium Voice Cloning, Training and Generation App - 1-Click to Install

Furkan Gözükara — Fri, 04 Sep 2026 15:28:31 +0000

IndexTTS 2.5 Premium - The Complete Beginner-to-Advanced Guide

App Download Link

You can get the app installer from here https://www.patreon.com/SECourses/posts/indextts-2-5-and-139297407

Quick Info

Voice cloning, long-form narration, caption-timed audio and MP4, batch production, dataset preparation, LoRA/DoRA training, checkpoint evaluation, listening grids, speaking-rate calibration, and low-VRAM operation - all in one tested workflow.

V6 maintenance update: the main workflows and registered settings remain compatible with the screenshots, while the following user-visible behavior is new or corrected.

Record a reference directly from the browser microphone, alongside upload, local-path, library, adapter, and recent-output sources.
Local media paths are staged safely for browser playback, and unsupported video codecs receive an FFmpeg-generated browser preview without changing the audio used for cloning.
Voice Generation now defaults to in-process model reuse; isolated subprocess mode remains available when hard cancellation and complete VRAM release matter more than repeat-run speed.
Candidate players, dataset reference players, feature-cache handoff, completed batch summaries, CPU diagnostics, and narrow-screen header controls now restore or update reliably.
A lazy-rendered Changelog tab now keeps the public v6.0-to-v4.0 release history and official SECourses project links inside the app.

Choose a route:

Beginner route: install, choose a reference, enter text, and generate a first voice.
Production route: control segmentation, pauses, subtitles, emotion, duration, formats, and batches.
Voice-training route: prepare clean data, cache features, train LoRA or DoRA, evaluate checkpoints, and deploy the best adapter.
Hardware route: choose a VRAM tier, BF16 or INT8, block swapping, auxiliary-model residency, and verify the setup with the isolated benchmark.

1. Install and Start the App

Windows one-click install or update

Place the downloaded package in a short writable path such as G:\Index_TTS_v4. Before the first install, provide Python 3.12.10 or newer, Git, FFmpeg, CUDA 13.0, cuDNN 9.17 or newer, and Visual Studio Community with the C++ workload. Then run Windows_Install_or_Update.bat.

The installer clones or updates Premium_IndexTTS2_SECourses, creates its Python 3.12 virtual environment, installs uv, resolves the requirements, and runs the model downloader. If it detects local source changes, it deliberately skips git pull rather than overwriting them.

Run Windows_Install_or_Update.bat once for a fresh install and again whenever an update is announced.
Run Windows_Start_App.bat for normal use. It activates the environment, configures the CUDA allocator and model cache, and launches the Gradio UI.
Wait for the local URL, normally http://127.0.0.1:7860, and open it in a browser.
For a specific GPU, add set CUDA_VISIBLE_DEVICES=1 to the start BAT or launch python webui.py --device cuda:1 from the activated environment.

Useful launch flags are --port, --host, --share, --model_dir, --verbose, --no-browser, and --device. For example:

Windows_Start_App.bat --port 7861 --device cuda:0 --no-browser

Linux, RunPod, SimplePod, and Massed Compute

Use the supplied RunPod_Install_IndexTTS.sh or Massed_Compute_Install.sh, then follow the matching instruction text file in the package. The current Linux installers require root or sudo, install aria2, xz, Git LFS, and FFmpeg/ffprobe n9.0, provision a current Python 3.12 environment, install the application requirements, and download the IndexTTS 2.5 models. The Linux launcher then enters the app folder, activates venv/bin/activate, clears a conflicting LD_LIBRARY_PATH, and starts Gradio with --share. Keep the generated public Gradio link private because anyone who has it can use the running GPU.

bash Linux_Start_App.sh --device cuda:0

If startup reports missing model files, run Windows_Model_Download_and_Fix.bat on Windows or use Models & Performance > Download / verify base models after the interface opens.

2. Know the Workspace Before Generating

The header controls reusable settings; the tab row separates generation, batch work, dataset preparation, training, checkpoint comparison, performance, help, and release history. A fresh V6 install selects the quality preset automatically; an updated installation restores its last-used preset. Start in Voice Generation and keep the quality preset for the first successful output.

Figure 1. The first screen is a working console, not a landing page: reference and text are on the left and center, while run controls, progress, results, and logs stay on the right. **Open / close all sections* is the fastest way to expose or collapse advanced controls.*

At startup, earlier result panels stay clean. Load last values restores the most recently saved values across every tab. A system preset is marked with a star and is read-only; a user preset can be created, overwritten, loaded, or deleted. Reset returns the registered controls to the system default preset without deleting model files or generated outputs.

Voice Generation: one script, one reference workflow, optional candidates and media output.
Batch Generation: many TXT, SRT, VTT, or SBV jobs with shared or per-file references.
LoRA Dataset Preparation: turn raw audio/video/captions into quality-controlled training clips.
LoRA / DoRA Training: train, validate, sample, save, resume, analyze, and manage adapters.
Checkpoint Grid: compare base, recommended, final, and epoch checkpoints with identical inputs.
Models & Performance: fit the model to the GPU, verify files, and benchmark real VRAM use.
Help: embedded quick starts, parameter guidance, pause syntax, and recovery steps.
Changelog: read newest-first release notes and open the official Patreon or GitHub project pages.

3. Make the First Voice Clone

Choose the speaker and write the script

Use a clean, single-speaker reference with little music, reverb, or room noise. Eight to fifteen seconds is a strong starting point. Audio, video, and direct microphone recording are accepted; video is decoded through FFmpeg and the audio preview shows the exact material generation will use.

Figure 2. Upload or drop the reference, type the script, choose its language, and press **Generate voice. V6 also places a **Record from microphone* accordion directly beneath the upload surface. The right column exposes cancellation, progress, current item, speed, ETA, the generated player, and the last 60 log lines.*

Drop an audio or video file into Reference Voice, expand Record from microphone and record a WAV in the browser, or enter an existing local path in Reference media path and press Load path. Stop the microphone recording before generating; the recorded WAV becomes the active Reference Voice.
If only part of a long source is clean, enter ranges such as 1:4;7.5:12 or 01:02-01:08, then press Extract ranges. Ranges are joined in the order written.
Type the words to synthesize. Match Language to the script, not necessarily to the reference speaker.
Keep Max tokens per segment on Auto or the language-aware default for the first run.
Press Generate voice and do not close the terminal while the worker is active.

The Reference audio library scans reference_audios. Refresh rescans it and Load path applies the selected entry. When no manual reference is present, an enabled LoRA/DoRA can supply its recommended reference automatically; otherwise the newest compatible library file is the final fallback. Clear removes uploads, recordings, previews, the path field, and the current manual selection.

For a local path outside the app's normal output, dataset, adapter, reference, or temporary folders, V6 stages a browser-safe hard link or copy under .ui_state/reference_media. A video whose codec cannot play directly in the browser receives a cached preview under .ui_state/reference_video_previews; generation still uses the audio extracted from the original source.

Understand automatic text segmentation

Long input is split before synthesis so each section stays inside the model and VRAM budget. The live preview shows section number, type, text or pause, and details before you commit GPU time.

Figure 3. Shorter segments are safer on low VRAM and make failures easier to retry; longer segments can preserve phrasing but reserve more semantic and diffusion memory. The non-CJK budget scale leaves room for English, Arabic, and Spanish subword expansion.

Use punctuation to create natural boundaries. Enable pause tags when you need exact silence: [pause:500ms], [pause:0.8s], and <pause=0.5> are accepted forms. A pause becomes its own section and is inserted without asking the speech model to invent silence.

Text normalization expands text before phonetic processing and is recommended for ordinary numbers and punctuation. Disable it only when you have a deliberate pre-normalized script and have checked pronunciation. If the WeText backend rejects an unusual English or Chinese fragment, V6 keeps the original text and logs a concise warning instead of dropping the fragment.

Use captions and create an MP4

Upload SRT, VTT, or SBV captions to replace or organize the script. With Use caption cue timing, each caption unit is retimed to its cue slot and cue start times are preserved. Add a still image only when an MP4 is required.

Figure 4. Caption timing is ideal for localization, dubbing drafts, and slide narration. Without the timing checkbox, caption text is still synthesized but flows naturally; without a still image, the job remains audio-only.

Plain captions: use subtitle text as a structured script while allowing natural timing.
Cue-timed captions: preserve subtitle starts and fit each unit to its slot.
Still-image video: combine the completed audio and uploaded image into an MP4.
Caption plus pause tags: pause parsing still occurs inside caption text when enabled.

4. Shape Emotion, Sampling, and Timing

Four emotion sources

Emotion is independent from speaker identity. Select Same as speaker for the simplest clone, Emotion reference audio to transfer delivery from another clip, Manual vectors for direct channel control, or Emotion text to describe the desired performance.

Figure 5. **Emotion weight* blends the chosen emotion source with the speaker tone. Manual mode exposes joy, anger, sadness, fear, disgust, depression, surprise, and calm; tuned biases and the maximum vector sum keep a combination from becoming unnaturally extreme.*

For emotion-reference work, use a clip whose delivery is clear even if the voice is different: the speaker reference supplies identity and the emotion reference supplies style. Random emotion exemplar varies the internal exemplar used with manual vectors. For reproducible comparisons, turn it off and keep the generation seed fixed.

Emotion-text mode needs the Qwen emotion model enabled in Models & Performance. Enter a short direction such as calm, reassuring, and quietly optimistic. If the field is blank, the app analyzes the speech text itself.

Autoregressive and diffusion controls

The autoregressive stage decides semantic tokens; the diffusion stage turns those tokens and conditioning into acoustic detail. Change one family at a time so you can hear what caused the difference.

Figure 6. The quality preset uses sampling, temperature 0.8, top-p 0.8, four beams, CFM temperature 0.9, and 40 diffusion steps. The system default preset uses three beams, CFM temperature 1.0, and 25 steps. A fixed seed plus a single candidate is the cleanest diagnostic setup.

Temperature: lower is steadier; higher is more varied but can be less stable.
Top-p and top-k: restrict the token candidate pool. Top-k 0 disables that filter.
Beams: can improve stability, but multiplies time and VRAM. Optional acceleration expects beams 1.
Repetition penalty: prevents semantic-token loops; keep the established default unless diagnosing repeats.
Length penalty: affects beam search only; 0 is neutral.
Max mel tokens: a safety ceiling, not a requested duration.
Candidates: consecutive seeded alternatives from one request; each costs another generation.
Diffusion steps: 12-16 is a faster draft range, 25 is the registered/system-default value, and the quality preset uses 40; 35-50 can refine difficult material.
CFG rate and CFM temperature: control conditioning strength and diffusion variation.
CFM cache length: lower it only when reserved VRAM is the problem.

Pacing, silence, and reference processing

Use Speaking rate for the voice's pace, Section silence for joins, and Target duration only when the entire assembled output must approach a known length. A trained voice can store a measured speaking rate and apply it automatically.

Figure 7. Reference limits normally stay at 15 seconds. Semantic layer 17 is the trained recommendation. Reusing speaker conditioning for emotion is the fast default when there is no separate emotion source.

Off: no whole-output duration target.
Natural: regenerate timing toward the target instead of mechanically editing the finished waveform.
Pad: append silence only when output is shorter; it never speeds up or truncates longer speech.
Trim: cut a longer assembled result to the exact target.
Max consecutive silence tokens: 0 disables token trimming; use it only to suppress unusual model silences.
Latent multiplier: the natural-duration factor passed to the engine; leave 1.72 unless running a controlled timing experiment.

Formats, audio finishing, and execution mode

Give the task an optional safe filename, keep the used reference for reproducibility, and choose WAV, MP3, or still-image MP4 behavior. Bypass preserves model audio; the other tuning presets use FFmpeg post-processing with optional explicit overrides.

Figure 8. V6 system presets start with in-process reuse for faster repeat runs. Enable isolated subprocess mode when Cancel must terminate the complete model process and release its VRAM after the job; isolation trades away resident-model reuse.

MP3 bitrate 256k is a strong voice quality and file-size balance; WAV candidates always remain available.
Edge-silence trimming removes only sufficiently long leading or trailing silence.
Low cut, high cut, gain, LUFS, and de-ess values override the chosen tuning preset when supplied.
Every non-bypass tuning preset preserves the source sample rate in V6.
In-process reuse is the V6 default; isolated subprocess is an intentional per-preset opt-in.
Section batch size 1 is safest. Raise it only within the active VRAM tier hint.
Low-memory mode chooses sequential, aggressive-memory paths.
Prevent VRAM accumulation clears autoregressive caches between sections; it is slower but useful for long books.
Verbose logging exposes model inputs, timing, and detailed diagnostics in the terminal and live log.

5. Run, Cancel, Review, and Reuse Outputs

During a run, watch item number, elapsed time, ETA, speed, current stage, and the live log. Cancel stops new work and, in subprocess mode, can terminate the worker cleanly instead of leaving GPU memory occupied.

Figure 9. A successful run ends with a playable result and a final summary. **Open outputs folder* jumps to the task directory containing audio, media, reference copies, metadata, and logs selected by the output settings.*

If generation seems stuck, read the last log line before cancelling. First-time model loads and optional compilation can be much slower than later runs. If cancellation leaves a model resident in reuse mode, use Models & Performance > Unload model / free VRAM.

When Candidates is above one, listen to every player before choosing. The displayed seed lets you repeat or compare a run with the same stochastic starting point; settings and output metadata explain exactly what created each file.

Figure 10. **Recent outputs* lists the last ten generated tasks. Select a good result and press Load selected output into reference to perform iterative voice cloning, continuation work, or a clean second-generation reference test.*

6. Batch Generation for Many Scripts

Build a queue from uploaded TXT/caption files, pasted text, or a local folder. The batch uses the shared Voice Generation settings, so first prove one representative item in the single-generation tab.

Figure 11. Naming supports {index}, {name}, and {stem}. Keep output in a safe subfolder under outputs, choose shared or same-stem per-file references, then choose cancellable subprocess execution or faster in-process reuse.

Load mixed TXT, SRT, VTT, and SBV sources or point the folder field at a local collection.
Choose Shared reference to use the active Voice Generation reference for every item.
Choose Per-file reference when each text/caption file has a same-stem audio file beside it.
Enable Continue after item errors for unattended queues; failed rows are recorded while later items continue.
Press Generate batch, monitor the table and log, and use Open batch folder when complete.

The tested queue mixed a plain TXT item with an SRT item and completed both. Caption timing is now applied per item, so a subtitle entry can use its cues without incorrectly forcing the TXT entry through a caption-only path.

Figure 12. Confirm that every requested item has a completed row, playable output, path, elapsed time, and no hidden error in the log. **Cancel batch* is safe at an item boundary and terminates the active subprocess when needed.*

7. Prepare a High-Quality Voice Dataset

Inputs, names, and scanning

Dataset preparation accepts media files, folders, metadata.csv, or already segmented WAV+TXT folders. Use one safe dataset name and confirm the output root before scanning, especially when Overwrite dataset will later be enabled.

Figure 13. Recursion discovers nested media. A fixed speaker name labels every segment; **Speaker from folder* instead uses each source parent folder, which is useful for a multi-speaker collection.*

Press Scan inputs before processing. The discovered-media table and statistics expose supported files, sidecar captions, durations, and warnings without spending time on transcription or cutting.

Figure 14. Fix missing paths, unsupported extensions, or accidental extra speakers at this stage. **Refresh* updates existing datasets, and Open dataset folder inspects a selected prepared dataset.*

Transcripts and sentence-aligned segmentation

Prefer sidecars uses SRT/VTT/SBV text and timing when available and lets Whisper fill missing alignment. Sentence-aligned mode with Whisper word times is the recommended CUDA workflow because it preserves complete phrases instead of arbitrary waveform chunks.

Figure 15. Remove bracket notes such as [music], deduplicate rolling captions, and drop repeated spoken sentences. The measured quality window is 4 to 20 seconds with a 14-second target; edge padding and silence snapping avoid clipped consonants.

Sentence boundaries require punctuation and are the cleanest option.
Sentence-or-pause recovers more material at aligned-word pauses, with a small risk of less natural cuts.
Maximum cue gap controls when nearby caption cues may merge.
Minimum pause boundary controls how much silence is needed when punctuation does not provide an edge.
Minimum and maximum word counts remove fragments and implausibly dense transcripts.

Cleanup and objective quality gates

Trim leading and trailing silence before filtering, normalize loudness for consistent gradients, and retain the required 24 kHz sample rate. The remaining gates reject weak alignment, impossible speaking rates, very quiet audio, clipping, and excessive silence.

Figure 16. Defaults are conservative voice-training safeguards: -20 LUFS leaves headroom, 40 dB is a gentle trim threshold, and a 0.001 clipping ratio permits at most 0.1 percent clipped samples. Leave maximum silence ratio blank to disable that optional whole-segment filter.

Do not loosen several gates at once. Review rejected counts and playable segments, then change the single rule that is demonstrably excluding good speech. Bad transcripts or clipped recordings cannot be repaired by more training epochs.

Destination, references, and deterministic smoke tests

Export reference candidates so training samples, validation, and later Voice Generation can reuse the cleanest clips. Maximum segments set to 0 processes everything; a small number is ideal for validating paths and settings before a long run.

Figure 17. The preparation seed makes candidate ranking and randomized work reproducible. **Overwrite dataset* replaces the same named directory, so leave it off until the scan and destination are unquestionably correct.*

Press Prepare dataset and follow item progress, speed, ETA, stats, and the log. A successful summary reports accepted and rejected segments, total duration, manifests, reference candidates, and the output path.

Figure 18. Completion means the files were produced, not that every segment is automatically good. Use the inspection panel next; **Cancel* is available throughout long extraction, alignment, and segmentation jobs.*

Sort through the prepared-segments table, inspect durations and warnings, and click representative rows to play the exact training waveform. Look for clipped words, wrong captions, background speakers, music, and long silence.

Figure 19. The waveform and duration distribution make outliers visible. A smaller, clean, correctly transcribed dataset normally beats a larger noisy one for a single voice.

Precompute the feature cache

Feature caching runs the expensive base-model preprocessing once and lets training consume compact cached samples. Choose the intended model directory and device, then press Cache features now.

Figure 20. Monitor cache progress and failures just like preparation. If a cached sample exceeds text or semantic-code limits, correct or remove the source segment rather than hiding the failure.

When caching completes in V6, the prepared dataset remains selected, its cached status refreshes immediately, and the Training tab receives the same dataset automatically. You no longer need to reselect it just to make the training summary notice the new cache.

8. Train a LoRA or DoRA Voice

Adapter architecture

Select a prepared cached dataset, enter a safe adapter name, and choose LoRA or DoRA. DoRA is the measured quality default; LoRA uses slightly less compute. The recommended measured setup is rank 128 with alpha 129.

Figure 21. Attention and MLP targets are recommended for voice fidelity. Speaker projection is a small fully trained module; emotion layers and the mel embedding/head are advanced additions that expand the trainable surface.

Start with the default trainable modules. Increasing rank or enabling extra modules does not guarantee better speech; it increases file size, compute, and overfitting risk. The app inspects type, rank, and alpha when resuming so incompatible weights are rejected early.

Optimizer, schedule, and effective updates

The measured batch-1 baseline uses AdamW, cosine decay, learning rate 4e-5, 200 warmup steps, 10 epochs, batch size 1, and accumulation 1. With that setup, every training clip produces one optimizer update per epoch.

Figure 22. max_steps=0 derives the plan from epochs; use a tiny nonzero maximum for a smoke test. Gradient clipping limits spikes, while label smoothing and stronger weight decay should be changed only in response to measured validation behavior.

Fused AdamW can be faster on compatible CUDA builds; ordinary AdamW is the portable choice.
Betas 0.9,0.99 and the default epsilon are established stable optimizer values.
Mel loss is the main acoustic-token objective; text loss is auxiliary.
Speaker reference other and emotion reference follow_speaker imitate normal inference better than self-conditioning.

Validation and early stopping

Keep a validation split so the app measures unseen clips rather than only memorization. Five percent is the general default; tiny test datasets may need a larger fraction simply to contain at least one validation item.

Figure 23. Validation can run every N steps and/or at epoch boundaries, with a batch cap for large datasets. Early stopping requires both patience and a minimum improvement; patience 0 disables it.

Validation reference mode other uses a different same-speaker clip for both vectors and most closely matches real cloning. The training smoke test used four segments with a 25 percent split, producing three training items and one unseen validation item. If a tiny dataset produces zero validation items, V6 safely skips automatic measured checkpoint evaluation and records the reason in the log instead of starting an invalid evaluation job.

Precision and low-VRAM training

BF16 base weights and BF16 mixed precision are recommended on modern NVIDIA cards. Gradient checkpointing trades extra computation for much lower activation memory and is required when training with GPT block swapping.

Figure 24. INT8 ConvRot reduces frozen base-weight memory. Block swapping streams selected frozen GPT blocks through a CPU ring; pinned memory improves transfer speed, while ring size 1 uses the least VRAM.

Use Apply VRAM tier defaults instead of guessing a low-memory combination. If training still OOMs, reduce batch size first, keep gradient checkpointing enabled, add block swap, and close other GPU applications. Do not compensate by raising gradient accumulation unless you intentionally want a different optimization regime.

Saving, resume modes, and automatic evaluation

Save an epoch checkpoint so the best-sounding voice is not forced to be the final epoch. Keep-last 0 retains every epoch; Save best also tracks the lowest validation loss. BF16 adapter files are smaller, while FP32 preserves full update precision.

Figure 25. **Weights only* loads an adapter into a fresh schedule at step 0. Continue run restores optimizer, scheduler, step, RNG, and data position when train state exists. Automatic analysis is CPU-only; measured checkpoint evaluation starts only after the training model releases GPU memory.*

Save train state for exact continuation from best, final, or interrupted checkpoints.
Save train state with every epoch only when you need to continue from an arbitrary epoch; it costs substantially more disk.
Automatic evaluation may include the base model, a deterministic training subset, selected strengths, and a timeout that does not invalidate completed training.
The recommended checkpoint is written into machine-readable and plain-language analysis files.

Per-epoch listening samples

Enable training samples to hear progress at a fixed epoch interval. Keep one representative short sentence, one reference, one language, and one seed so differences come from the checkpoint rather than changing inputs.

Figure 26. The isolated sampling worker uses its own runtime tier, free-VRAM threshold, and timeout. Blank reference selects the dataset's best candidate automatically; Auto language follows the prepared dataset.

Training samples mirror the important Voice Generation parameters: beams, temperature, top-p, top-k, repetition, emotion weight, diffusion steps, CFG, text-token ceiling, length penalty, mel-token ceiling, and speaking rate.

Figure 27. Lock these values before training if you intend to compare epochs by ear. Otherwise a changed decoding setting can sound like a training improvement or regression.

Paths, seed, workers, and diagnostics

The final group sets output root, model directory and YAML, training device, attention backend, deterministic seed, data workers, and log frequency. Relative paths resolve from the app directory.

Figure 28. Two workers is a safe Windows/Linux default; use zero when debugging loader problems. Log every step for a short run and less frequently only when an extremely fast long run makes logging expensive.

Start, stop, and read the live dashboard

Press Start training only after the optimizer-update plan looks sensible. The live dashboard reports epoch, optimizer step, loss, validation, learning rate, gradient norm, throughput, elapsed time, ETA, VRAM, samples, and the worker log.

Figure 29. **Stop* requests a graceful checkpointed stop. Force stop terminates a nonresponsive worker. Both paths were exercised; use force only after the normal stop path has had time to save.*

The real smoke run completed two optimizer steps on a DoRA rank-8 adapter, with a three-item train split and one-item unseen validation split. The final dashboard exposed its checkpoint, sample, measured values, and analysis actions.

Figure 30. Use **Open output folder* for artifacts, Compare in grid for controlled listening, and Use best checkpoint to send the recommendation to Voice Generation. Completion alone is not a reason to choose the final epoch.*

The completed run remains selected while analysis compares training and unseen loss. Read the recommendation, sustained-overfitting marker, and checkpoint table before deploying an adapter.

Figure 31. The handoff buttons avoid manual paths: compare the run in Checkpoint Grid or use its recommended checkpoint directly in generation.

The manager inventories adapter type, rank, alpha, steps, dataset, recommendation, speaking rate, path, and health. Refresh rescans, Open folder inspects files, and Delete removes only after the UI confirmation flow.

Figure 32. Use the manager to distinguish a final adapter from epoch checkpoints and interrupted runs. A valid entry should have metadata and a readable safetensors file before it appears in Voice Generation.

9. Analyze and Compare Checkpoints

Generalization analysis and measured evaluation

Select one adapter folder and press Analyze training log. The CPU-only analysis compares train and validation trends, identifies the lowest unseen loss, and detects a sustained overfitting region rather than assuming the newest file is best.

Figure 33. A recommendation based on validation is a shortlist, not the final listening decision. Use the next measured and audible comparisons with the same references and text.

The checkpoint table combines epoch/step identity, paths, validation values, verdicts, and the recommended marker. Base model provides a reference-only baseline when enabled.

Figure 34. **Use best checkpoint* updates the Voice Generation adapter choice. Evaluate checkpoints now starts an isolated inference-like measurement over the configured validation and training subsets.*

Evaluation references can match training validation or use a different clip of the same speaker, closer to ordinary voice cloning. The job fixes its examples and settings across every checkpoint.

Figure 35. Live progress shows which row is loading and scoring. The test executed a real three-row comparison, not a mocked UI state.

The completed measurement reports comparable scores and a best-checkpoint recommendation. Inspect failed rows and reference selection before trusting a ranking.

Figure 36. Use measured evaluation to narrow the field, then rely on the listening grid for intelligibility, identity, pacing, emotion, noise, and artifacts that one scalar cannot fully express.

Create a fair listening grid

Choose base, recommended, final, and selected epoch checkpoints; add strengths from 0 to 4; then provide one text per line and one reference path per line. Every cell receives the same language, text segmentation, reference, and seed.

Figure 37. **Use LoRA / DoRA reference* inserts the stored reference. Add dataset reference candidates adds clean prepared clips. The runtime estimate grows with checkpoints x strengths x texts x references.*

Lock temperature, top-p, top-k, beams, repetition penalty, diffusion steps, CFG, segment budget, emotion weight, and speaking rate before running the grid.

Figure 38. Seed -1 draws one random seed at grid start and then reuses it for every cell. Enter an explicit seed when a grid must be recreated later.

Press Generate grid and watch cell-by-cell progress. The worker saves state under outputs/grids, so the log, table, audio, and metadata stay together.

Figure 39. **Cancel* stops the active grid worker; completed cells remain useful evidence. Do not change runtime/model settings in another tab while an isolated comparison is loading.*

Completion summarizes every generated cell, its checkpoint, strength, text/reference index, seed, duration, and path. Use Open grid folder to inspect the saved Markdown, metrics, log, and audio files.

Figure 40. A complete table verifies coverage; the audible result view is where the final voice decision happens.

Listen horizontally with the same text/reference and vertically across base, best, final, epochs, or strengths. The saved production grid was reloaded successfully and displayed all 12 audio players.

Figure 41. Rate each row for identity, pronunciation, prosody, background noise, breath artifacts, repeated syllables, and consistency. Prefer the earliest clean checkpoint when later gains are negligible and overfitting begins.

Calibrate the voice's real speaking pace

Press Calibrate speaking rate from this grid after generating representative audio. The app compares generated words per second with the voice's training recordings and saves a per-adapter multiplier.

Figure 42. Voice Generation can automatically apply the stored value when that adapter is selected. The production calibration measured about 0.944, while a separate smoke calibration proved the full write-and-reload path.

10. Fit Runtime to the GPU

Inventory, tiers, and applying settings

Refresh GPU inventory, select the actual device, and choose Auto or a 6, 8, 10, 12, 16, 24, or 32 GB tier. Every tier is a coordinated set of model, swapping, cache, and batch hints designed to retain roughly 2 GB of headroom.

Figure 43. **Apply runtime* commits the current model and memory settings. Changing a dropdown alone does not rebuild an already loaded model; unload before a major variant or device switch.*

BF16, INT8, attention, and acceleration

BF16 is the official quality path on modern NVIDIA GPUs. INT8 ConvRot reduces GPT weight memory while retaining the required IndexTTS 2.5 codec and vocoder stack. FP32 is primarily a CPU-compatible fallback.

Figure 44. SDPA is the compatible attention default. FlashAttention 2 needs its optional package. The acceleration engine uses the fast CUDA-graph/flash path and expects beams 1; torch compilation trades a slower first run for repeated-workload speed.

Enable the Qwen emotion model for Emotion text mode; on-demand residency avoids eager startup cost.
BigVGAN CUDA kernel uses the optional fused activation kernel when available.
CFM estimator BF16 autocast saves activation VRAM and can slightly change output.
DeepSpeed is a legacy optional loader and should stay off unless installed intentionally.

GPT block swapping

Zero keeps every GPT block resident. Minus one lets the runtime fit automatically; positive values up to 24 stream that many blocks from CPU RAM. More swapping reduces VRAM but increases transfer overhead.

Figure 45. Ring size 2 overlaps copies and compute; ring size 1 minimizes memory. Keep pinned swap memory on for faster CPU-to-GPU transfer unless the host cannot lock enough RAM.

Runtime CFM cache length sets the upper reservation, VRAM reserve protects against allocator spikes, and the section-batch hint informs Voice Generation. Treat the hint as an advisory maximum, not a promise when another process is using the GPU.

Selecting cpu now shows CPU diagnostics mode rather than a misleading GPU fit estimate. GPU VRAM limits do not apply there, and FP32 is the compatibility-oriented choice; CPU synthesis is primarily a diagnostic fallback and is much slower than CUDA.

Place auxiliary models independently

Semantic model, Qwen emotion, CampPlus, semantic codec, S2Mel, and BigVGAN can each stay on GPU, stay on CPU, or load on demand according to the choices provided by the active tier.

Figure 46. Keep frequently used synthesis stages resident when VRAM allows. Move optional or infrequent models off GPU first when reducing idle footprint.

Verify and download model files

The model table checks required base and optional INT8 files. Download / verify base models repairs the official stack; Download INT8 model fetches the memory-efficient GPT; Refresh file status rechecks disk; Open model folder opens the configured directory.

Figure 47. The INT8 path now has a verified fallback: if the hosted artifact is unavailable, the downloader can convert the local BF16 model instead of presenting a feature that cannot be obtained.

Run the isolated VRAM benchmark

Select a tier and press Run VRAM benchmark when the chosen GPU is otherwise idle. The UI uses a short deterministic calibration - beams 1, 60 text tokens, batch 1 - while the standalone CLI can still run a heavier stress profile.

Figure 48. The benchmark uses a bundled reference voice, runs in an isolated subprocess, and checks whether the selected tier truly fits rather than estimating from file size alone.

Watch the subprocess log for model load, warmup, synthesis, peak allocation, and fit status. On Windows WDDM, the idle gate allows the larger of 1 GB or 10 percent of physical VRAM so driver accounting does not falsely block an idle card.

Figure 49. A benchmark can be cancelled without contaminating the main app model because it owns a separate process.

The real 32 GB BF16 test fit successfully on the RTX 5090: 28.714 seconds of audio, real-time factor about 0.458, and peak allocated VRAM about 6.212 GB for the short calibration workload.

Figure 50. Read **fit, peak VRAM, RTF, and any error together. Apply the tier after a pass, then use **Unload model / free VRAM* whenever you want to release the in-process runtime.*

11. Built-in Help and Recovery

The Help tab repeats the shortest successful path for a first clone and explains the reference-only and trained-adapter workflows. Use it as an in-app checklist while this longer guide stays open separately.

Figure 51. The best diagnostic habit is to return to one clean reference, one short sentence, the quality preset, base model, and section batch size 1.

Workflow help covers speaking-rate calibration, checkpoint comparison, VRAM tiers, and the main parameter families. Follow a complete workflow rather than changing unrelated controls until the symptom disappears.

Figure 52. Presets make experiments repeatable: duplicate a known-good user preset and change one variable for each A/B comparison.

The final help area documents pause syntax, reference guidance, links, and recovery steps. Copy a supported pause form exactly and check the live section preview before generation.

Figure 53. For OOM: unload, lower batch size, choose a lower tier or INT8, add block swap, shorten sections, and rerun the benchmark. For bad speech: first inspect reference quality, language, text segmentation, seed, and adapter strength.

Read the V6 release history

V6 adds a lazy-rendered Changelog tab after Help. Open it to read the newest-first v6.0 through v4.0 release notes, including fixes that may affect an older workflow, and to reach the official SECourses Patreon and GitHub repository. The tab was added after the original V5 screenshot set, so it is documented here rather than shown in those captures.

12. Presets, Themes, and Repeatable Work

A universal preset stores every registered setting across every tab. Select a system preset for a protected baseline, or type a new name and press Save to create a user preset. Selecting an existing user preset allows an intentional overwrite.

Figure 54. **Load* applies a preset, Delete requires the confirmation flow and only removes user presets, Reset restores defaults, and Load last values recovers the most recent working state. Fresh installs select the quality system preset; upgrades retain the last-used preset. Unknown old keys are ignored and missing new keys receive defaults.*

Press Light / dark theme for the preferred presentation. Theme changes the interface only; it does not change audio, model precision, presets, or output files.

Figure 55. The 4K capture confirms the same dense workspace remains legible in light mode. Switch back at any time without restarting a job.

13. Deploy a Trained Voice and Use Advanced Controls

Load, scale, and merge an adapter

Choose an adapter in Voice Generation. Its summary exposes type, rank, alpha, steps, recommendation, and saved metadata. Strength 1.0 is the trained scale; lower values are subtler and higher values are stronger.

Figure 56. Auto-reference and auto-speaking-rate remove two common deployment mistakes. BF16 can temporarily merge the adapter into base weights for speed; the app restores base weights before switching, and INT8 stays on the unmerged path.

Emotion-reference mode

Select Emotion reference audio, upload a short expressive clip, and set emotion weight. Keep the identity reference separate: this mode is for transferring delivery, not replacing the speaker.

Figure 57. Use the maximum emotion-reference length to discard unneeded material and compare at a fixed seed before increasing the blend.

Manual-vector mode

Select Manual vectors and blend the eight channels. Start with one dominant channel, apply tuned biases, and keep the sum limit at 0.8 before attempting compound emotions.

Figure 58. A vector is continuous rather than a label: 0.2 calm plus 0.1 sadness is a different conditioning target from a full-strength sadness instruction.

Natural-language emotion mode

Select Emotion text and describe the intended performance in plain language. Keep it specific and short, such as restrained concern, then quiet relief, rather than writing a second script.

Figure 59. Qwen emotion must be enabled. The manual vector limits and biases remain available because the analyzed instruction ultimately becomes bounded emotion conditioning.

The tested app produced real outputs for emotion-reference, manual-vector, and emotion-text modes in addition to same-speaker emotion. Inspect the log to confirm which source loaded and listen against a neutral fixed-seed control.

Figure 60. A completed player, output path, duration, settings summary, and clean final status prove the selected emotion mode reached actual synthesis rather than only changing the UI.

Whole-output target duration

Enter a whole-output target only when the application has a real timing requirement. Natural changes synthesis timing, Pad only adds tail silence, Trim only cuts, and Off leaves the result unconstrained.

Figure 61. Combine duration targets cautiously with caption cue timing, speaking rate, section silence, and explicit pause tags because all of them influence the assembled timeline.

All three active paths were generated. A 3-second Natural target produced about 2.995 seconds; a 1.5-second Trim target produced exactly 1.500 seconds; a 4-second Pad target remained about 4.156 seconds because the spoken source was already longer, which is correct pad-only behavior.

Figure 62. Always read the final measured duration rather than assuming every mode forces equality. Choose Natural for adaptive timing, Trim for a hard upper bound, and Pad for a hard minimum without speeding or cutting speech.

14. Practical Recipes

Clean first clone

Select the quality system preset and Base model.
Use one clean 8-15 second speaker reference and one short script in the correct language.
Keep same-speaker emotion, natural rate 1.0, seed fixed, one candidate, and section batch 1.
Generate, inspect the log, and save the good result plus used reference.

Long audiobook or course narration

Calibrate the trained voice's speaking rate and enable automatic rate loading.
Use punctuation, pause tags, and a conservative per-segment token budget.
For long unattended jobs where complete post-job VRAM release matters, enable subprocess mode; also enable prevent-VRAM-accumulation when segment-to-segment cache growth is the problem.
Test one chapter, save a user preset, then run folders through Batch with deterministic naming and continue-on-error.

Subtitle-timed localization video

Upload the translated SRT/VTT/SBV and enable caption cue timing.
Choose Natural duration behavior for model-based timing or Trim/Pad only when mechanical bounds are acceptable.
Add a still image for MP4, generate one representative cue set, and check starts and final duration.
Batch the remaining caption files only after the representative item passes.

Train and deploy a new voice

Scan media and captions, prepare 4-20 second sentence-aligned 24 kHz clips, inspect them, and cache features.
Train the measured DoRA baseline with validation, samples, epoch checkpoints, train state, and automatic analysis.
Evaluate base plus saved checkpoints with inference-like references, then generate a fixed-seed listening grid.
Select the best audible checkpoint, calibrate speaking rate, and enable automatic reference/rate in Voice Generation.

Recover from low VRAM or OOM

Cancel the job and use Unload model / free VRAM.
Close other GPU workloads and select the matching conservative VRAM tier.
Use INT8, more block swap, ring 1, lower CFM cache, on-demand auxiliary models, and section batch 1.
Shorten text sections, keep beams 1, apply runtime, and pass the isolated benchmark before retrying.

Reproducible A/B test

Save a baseline user preset.
Fix reference, text, language, seed, checkpoint, strength, emotion source, and runtime.
Change exactly one control, generate one or more candidates, and record the output metadata.
Use a listening grid when comparing checkpoints or strengths so every cell is locked automatically.

15. Verified Behavior and Repairs Included in This Build

This tutorial was not produced from screenshots alone. The original audit completed real synthesis, batching, preparation, feature caching, training, checkpoint evaluation, listening grids, calibration, model verification, and benchmarking. The V6 maintenance audit rebuilt the live interface, checked the changed controls and Changelog in Chrome, and passed 288 tests, with 37 environment-specific cases skipped and 15 upstream PyTorch deprecation warnings.

Mixed TXT/SRT batches now apply caption timing only to caption items instead of crashing the plain-text item.
Batched inference no longer leaks sequential target-duration arguments into the batch engine.
Saved listening grids retain and render their audio players, and idle polling no longer clears a loaded result.
Dynamically loaded checkpoint selections now survive dependent events and timer polling instead of failing Gradio validation against an empty startup choice list.
INT8 download now has a local BF16-to-INT8 fallback when the hosted artifact is absent.
The Windows benchmark idle gate tolerates normal WDDM reservation, uses a bundled reference, and the UI runs a practical short calibration profile.
V6 restores microphone reference recording and makes arbitrary local audio/video paths browser-safe, including cached previews for codecs the browser cannot play directly.
Dynamic candidate and dataset-reference players render after reload, feature caching refreshes the training handoff, and completed batch summaries are no longer overwritten by a polling race.
Acceleration now honors disabled top-k/top-p limits, preserves stop tokens and compute dtype, and surfaces internal failures instead of silently returning an empty result.
Audio tuning preserves sample rate, text-normalization failures retain the original fragment, zero-item validation skips automatic evaluation cleanly, and CPU mode no longer claims a GPU VRAM fit.
The new Changelog tab renders only when opened and presents the public v6.0-to-v4.0 history plus official project links without slowing initial tab rendering.
Every final annotated image passed an exact 3840 x 2160 dimension gate and was individually uploaded to the dedicated Hugging Face discussion.
The repaired selectable copy source passed a complete Patreon paste: all 62 hosted images became full-width native image blocks with all 62 alt texts, and the headings, lists, links, and final paragraph were retained.

A feature is treated as complete only when its control exists, its workflow reaches a real result, its failure/cancel path is understandable, and the tutorial shows where the user acts and where proof appears.

16. Complete Action Reference

These are the non-setting actions and result surfaces a regular user will encounter. The numbered guide above shows each in context.

Header: Load last values; open/close all sections; switch theme; save, load, delete, or reset a universal preset.

Voice Generation: Upload/load/clear a reference; record a microphone reference; extract time ranges; refresh the reference library; auto-select token budget; upload captions and a still image; generate; cancel; open outputs; refresh adapters; play generated/candidate audio; inspect recent outputs; load a selected output as a new reference.

Batch Generation: Upload text/caption files; generate the queue; cancel; open the batch folder; read progress; play results; inspect the per-item table.

Dataset Preparation: Refresh/open existing datasets; upload media; scan; prepare; cancel; cache features; open output; inspect discovered media and stats; select and play prepared segments.

Training: Refresh datasets; apply tier defaults; refresh resume sources; start; graceful stop; force stop; open output; compare in grid; use best checkpoint; inspect charts/sample/checkpoints; refresh/open/delete adapter-manager entries.

Checkpoint Grid: Refresh adapters; analyze logs; evaluate checkpoints; use best; use stored reference; add dataset candidates; upload a reference; generate/cancel a grid; open saved grid; calibrate speaking rate; play every result cell.

Models & Performance: Refresh GPU inventory; apply runtime; unload/free VRAM; download INT8; download/verify base models; refresh file status; open model folder; run the isolated benchmark.

Help: Read the quick starts, workflow guidance, parameter glossary, pause syntax, troubleshooting steps, and launch arguments.

Changelog: Open the newest-first v6.0-to-v4.0 release history and follow the official Patreon or GitHub project links.

17. Every Registered Setting

The appendix below is generated directly from the app's preset registry. It covers all 253 registered controls, including current defaults, ranges, choices, and the help text shown by the UI. A preset stores these values across tabs.

Voice Generation - 70 settings

Language - generation.language. Language code used by text normalization and pronunciation. (default "EN"; choices "ZH", "EN", "JA", "AR", "ES")

Max tokens per segment - generation.max_text_tokens_per_segment. Per-language defaults are recommended; shorter segments use less VRAM. (default 60; minimum 20; maximum 300)

Use caption cue timing - generation.use_caption_timing. Retimes each caption unit to its cue slot and preserves cue start times. (default false)

Auto-load the LoRA / DoRA recommended reference audio - generation.auto_lora_reference. Loads the LoRA / DoRA's saved reference whenever no manual Reference Voice is selected. (default true)

Auto-apply the LoRA / DoRA calibrated speaking rate - generation.auto_lora_speaking_rate. Uses the selected voice's measured pace; selecting None resets speaking rate to 1.0. (default true)

Emotion source - generation.emotion_mode. Use the speaker tone, another reference, eight manual vectors, or emotion text analysis. (default "Same as speaker voice"; choices "Same as speaker voice", "Emotion reference audio", "Emotion vector", "Emotion text")

Emotion weight - generation.emotion_weight. 0 keeps more speaker emotion; 1 follows the selected emotion source fully. (default 0.65; minimum 0; maximum 1)

Random emotion exemplar - generation.emotion_random. Randomizes the internal exemplar used with manual emotion vectors. (default false)

Emotion description - generation.emotion_text. Used only in Emotion text mode; blank analyzes the speech text itself.

Joy - generation.emotion_joy. Manual joy strength. (default 0; minimum 0; maximum 1)

Anger - generation.emotion_anger. Manual anger strength. (default 0; minimum 0; maximum 1)

Sadness - generation.emotion_sad. Manual sadness strength. (default 0; minimum 0; maximum 1)

Fear - generation.emotion_fear. Manual fear strength. (default 0; minimum 0; maximum 1)

Disgust - generation.emotion_disgust. Manual disgust strength. (default 0; minimum 0; maximum 1)

Depression - generation.emotion_depression. Manual depression strength. (default 0; minimum 0; maximum 1)

Surprise - generation.emotion_surprise. Manual surprise strength. (default 0; minimum 0; maximum 1)

Calm - generation.emotion_calm. Manual calm strength. (default 0; minimum 0; maximum 1)

Apply tuned emotion biases - generation.apply_emotion_bias. Recommended balancing prevents several emotion channels from dominating. (default true)

Maximum vector sum - generation.max_emotion_sum. 0.8 is the model-tuned recommendation; larger values can sound exaggerated. (default 0.8; minimum 0.1; maximum 2)

Joy bias - generation.emotion_bias_joy. Multiplier applied before the vector sum limit. (default 0.9375; minimum 0.5; maximum 1.5)

Anger bias - generation.emotion_bias_anger. Multiplier applied before the vector sum limit. (default 0.875; minimum 0.5; maximum 1.5)

Sadness bias - generation.emotion_bias_sad. Multiplier applied before the vector sum limit. (default 1; minimum 0.5; maximum 1.5)

Fear bias - generation.emotion_bias_fear. Multiplier applied before the vector sum limit. (default 1; minimum 0.5; maximum 1.5)

Disgust bias - generation.emotion_bias_disgust. Multiplier applied before the vector sum limit. (default 0.9375; minimum 0.5; maximum 1.5)

Depression bias - generation.emotion_bias_depression. Multiplier applied before the vector sum limit. (default 0.9375; minimum 0.5; maximum 1.5)

Surprise bias - generation.emotion_bias_surprise. Multiplier applied before the vector sum limit. (default 0.6875; minimum 0.5; maximum 1.5)

Calm bias - generation.emotion_bias_calm. Multiplier applied before the vector sum limit. (default 0.5625; minimum 0.5; maximum 1.5)

Sample - generation.do_sample. Recommended for natural variation; disable for greedy/beam decoding. (default true)

Temperature - generation.temperature. 0.8 balances expressiveness and stability. (default 0.8; minimum 0.1; maximum 2)

Top-p - generation.top_p. Nucleus sampling threshold; 0.8 is recommended. (default 0.8; minimum 0; maximum 1)

Top-k - generation.top_k. Candidate token cutoff; 0 disables top-k filtering. (default 30; minimum 0; maximum 100)

Beams - generation.num_beams. More beams can improve stability but increase time and VRAM. (default 3; minimum 1; maximum 10)

Repetition penalty - generation.repetition_penalty. 10 is the established model default. (default 10; minimum 1; maximum 20)

Length penalty - generation.length_penalty. Only affects beam search; 0 is neutral. (default 0; minimum -2; maximum 2)

Max mel tokens - generation.max_mel_tokens. Upper limit on generated semantic tokens per section. (default 1500; minimum 50; maximum 1815)

Seed - generation.seed. -1 chooses a fresh random seed; reuse a shown seed for repeatability. (default -1; minimum -1; maximum 4294967295)

Candidates - generation.num_candidates. Generates consecutive seeded alternatives; each adds generation time. (default 1; minimum 1; maximum 8)

Diffusion steps - generation.diffusion_steps. 25 is the quality default; 12-16 is faster and 35-50 can refine difficult audio. (default 25; minimum 2; maximum 100)

CFG rate - generation.inference_cfg_rate. 0.7 is recommended; high values follow conditioning more aggressively. (default 0.7; minimum 0; maximum 2)

CFM temperature - generation.cfm_temperature. 1.0 is the best-quality default; lower values reduce diffusion variation. (default 1; minimum 0; maximum 2)

CFM cache length - generation.cfm_cache_length. 8192 fits typical sections; lower values reduce reserved VRAM. (default 8192; minimum 1024; maximum 32768)

Non-CJK token budget scale - generation.segment_budget_scale_non_cjk. 0.72 leaves room for subword expansion in English, Arabic, and Spanish. (default 0.72; minimum 0.3; maximum 1)

Section silence (ms) - generation.interval_silence. Silence inserted between generated text sections; cue timing overrides this to zero. (default 200; minimum 0; maximum 2000)

Max consecutive silence tokens - generation.max_consecutive_silence. 0 disables token trimming; use only to suppress unusually long model silences. (default 0; minimum 0; maximum 200)

Latent multiplier - generation.latent_multiplier. 1.72 is natural duration; the runner converts this to the engine duration factor. (default 1.72; minimum 0.5; maximum 3)

Speaking rate - generation.speaking_rate. 1.0 is the model's natural pace; below 1.0 speaks slower, above 1.0 faster. A trained LoRA / DoRA can carry a calibrated value that matches the speaker's real pace. (default 1; minimum 0.5; maximum 1.5)

Target duration (seconds) - generation.target_duration_s. Leave blank unless a whole-output duration target is needed. (default blank; minimum 0.1; maximum 3600)

Target duration mode - generation.target_duration_mode. Natural regenerates timing; pad/trim only adjust the assembled result. (default "off"; choices "off", "natural", "pad", "trim")

Enable pause tags - generation.enable_pause_tags. Parses inline pause tags before tokenization. (default true)

Text normalization - generation.text_normalization. Recommended: expands and normalizes text before phonetic processing. (default true)

Maximum speaker audio length (s) - generation.max_speaker_audio_length. 15 seconds preserves enough identity without wasting reference compute. (default 15; minimum 3; maximum 90)

Maximum emotion audio length (s) - generation.max_emotion_audio_length. 15 seconds is recommended for an emotion reference. (default 15; minimum 3; maximum 90)

Semantic layer - generation.semantic_layer. Layer 17 is trained and recommended; changing it alters reference embeddings. (default 17; minimum 1; maximum 24)

Reuse speaker conditioning for emotion - generation.reuse_spk_cond_for_emo. Faster default-emotion path; enable when no separate emotion source is used. (default false)

Output filename - generation.output_filename. Optional safe basename; task numbering is used when blank.

Save used reference - generation.save_used_audio. Copies the active Reference Voice into the task folder for reproducibility. (default false)

Save MP3 - generation.save_as_mp3. Converts the final output to MP3; WAV candidates remain available. (default false)

MP3 bitrate - generation.mp3_bitrate. 256k is a strong quality/size balance for voice. (default "256k"; choices "128k", "192k", "256k", "320k")

Audio tuning preset - generation.audio_tuning_preset. Bypass preserves model audio exactly; other presets use FFmpeg post-processing. (default "bypass"; choices "bypass", "voice_clarity", "clear_narration", "deharsh", "warm", "normalize")

Trim edge silence threshold (ms) - generation.trim_silence_ms_threshold. 0 disables trimming; only edge silence at least this long is removed. (default 0; minimum 0; maximum 3000)

Low cut (Hz) - generation.tuning_low_cut_hz. Optional high-pass cutoff; leave blank to use the preset. (default blank; minimum 20; maximum 500)

High cut (Hz) - generation.tuning_high_cut_hz. Optional low-pass cutoff; leave blank to use the preset. (default blank; minimum 1000; maximum 24000)

Gain (dB) - generation.tuning_gain_db. Optional final gain before limiting. (default blank; minimum -24; maximum 24)

Loudness target (LUFS) - generation.tuning_loudnorm_i. Optional integrated loudness normalization target. (default blank; minimum -30; maximum -5)

De-ess amount - generation.tuning_deess. Optional attenuation around sibilance frequencies. (default blank; minimum 0; maximum 12)

Use isolated subprocess - generation.use_subprocess. Recommended: cancellation can terminate the complete model process and release VRAM. (default false)

Section batch size - generation.section_batch_size. 1 is safest; use the active VRAM tier hint before increasing this. (default 1; minimum 1; maximum 16)

Low memory mode - generation.low_memory_mode. Uses sequential paths and aggressive memory behavior for constrained GPUs. (default false)

Prevent VRAM accumulation - generation.prevent_vram_accumulation. Clears autoregressive caches between segments; slower but useful for long jobs. (default false)

Verbose logging - generation.verbose. Prints detailed model inputs and timing diagnostics to console and the live log. (default false)

Batch Generation - 5 settings

Naming pattern - batch.naming_pattern. Supports {index}, {name}, and {stem}; the extension is selected by Output settings. (default "{index:03d}_{name}")

Output subfolder - batch.output_subfolder. A safe subfolder below outputs; task folders are created inside it. (default "batch")

Reference mode - batch.reference_mode. Per-file mode looks for a same-stem audio file beside each text/caption file. (default "One reference for all"; choices "One reference for all", "Per-file reference")

Batch execution - batch.execution. Subprocess is easiest to cancel; reuse is fastest after the first in-process load. (default "Subprocess per item"; choices "Subprocess per item", "Reuse loaded model between items", "Reload in-process model per item")

Continue after item errors - batch.continue_errors. Records a failed row and proceeds to the next item instead of ending the batch. (default true)

Dataset Preparation - 45 settings

Input files or folders - dataset.inputs. Accepts media files, folders, metadata.csv, or pre-segmented WAV+TXT folders.

Dataset name - dataset.name. A single safe directory name created below Output root. (default "voice_dataset")

Output root - dataset.output_root. Dataset parent directory; relative paths resolve from the application folder. (default "datasets")

Scan recursively - dataset.recursive. Includes supported media in nested input folders. (default true)

Language - dataset.language. Language stored in manifest rows and used for Whisper. (default "EN"; choices "ZH", "EN", "JA", "AR", "ES")

Speaker name - dataset.speaker_name. Optional fixed speaker label stored in each segment.

Speaker from folder - dataset.speaker_from_folder. Uses each source parent folder name as the speaker label. (default false)

Transcript policy - dataset.subtitle_policy. Prefer sidecars is recommended; Whisper fills missing timing/text. (default "prefer_sidecar"; choices "prefer_sidecar", "whisper_only", "sidecar_only")

Whisper model - dataset.whisper_model. Hugging Face model id or local path used when transcription/alignment is needed. (default "openai/whisper-large-v3-turbo")

Whisper device - dataset.whisper_device. CUDA device recommended for sentence alignment; CPU works but is much slower. (default "cuda:0")

Segmentation mode - dataset.segmentation_mode. Sentence aligned uses Whisper word times with caption sentences and is recommended on CUDA. (default "sentence_aligned"; choices "auto", "sentence_aligned", "cue_boundaries", "whisper_only")

Force Whisper alignment - dataset.align_with_whisper. Compatibility alias that forces sentence_aligned mode. (default false)

Remove bracket annotations - dataset.remove_bracket_annotations. Removes caption notes such as [music] and [applause]. (default true)

Deduplicate rolling captions - dataset.dedupe_rolling_captions. Removes repeated text from live/rolling subtitle cues. (default true)

Drop duplicate sentences - dataset.drop_duplicate_sentences. Keeps one copy (best aligned) of every sentence that is spoken more than once, e.g. repeated intros or outros; recommended for voice training. (default true)

Target seconds - dataset.target_s. 14 seconds packs whole sentences into inference-length clips; measured best with a 20 second maximum. (default 14; minimum 1; maximum 30)

Minimum seconds - dataset.min_s. 4 seconds keeps enough voice context while retaining the measured quality range. (default 4; minimum 0.5; maximum 15)

Maximum seconds - dataset.max_s. 20 seconds covers 12-15-second inference segments and retains more source audio; 30 seconds measured worse. (default 20; minimum 2; maximum 40)

Maximum cue gap (ms) - dataset.max_gap_ms. Cues closer than this can merge into one sentence segment. (default 700; minimum 0; maximum 3000)

Edge padding (ms) - dataset.pad_ms. Small context padding avoids clipped consonants. (default 60; minimum 0; maximum 500)

Snap to silence - dataset.snap_to_silence. Moves segment boundaries toward nearby low-energy points. (default true)

Silence snap window (ms) - dataset.snap_window_ms. Search radius around a proposed boundary. (default 200; minimum 0; maximum 1000)

Minimum words - dataset.min_words. Drops fragments with too little transcript context. (default 2; minimum 0; maximum 30)

Maximum words - dataset.max_words. Drops transcript segments that are implausibly dense. (default 80; minimum 10; maximum 200)

Segment boundaries - dataset.boundary_mode. Sentence requires punctuation; sentence_or_pause recovers more audio at aligned-word pauses with a small risk of less natural cuts. (default "sentence"; choices "sentence", "sentence_or_pause")

Minimum pause for a boundary (ms) - dataset.min_pause_boundary_ms. A fragment edge must have at least this much silence when punctuation does not provide the boundary. (default 400; minimum 0; maximum 60000)

Trim silence - dataset.trim_silence. Trims leading/trailing low-energy audio before filtering. (default true)

Trim top dB - dataset.trim_top_db. 40 dB is a conservative silence threshold. (default 40; minimum 10; maximum 80)

Normalize loudness - dataset.loudness_normalize. Recommended for consistent training gradients across sources. (default true)

Target LUFS - dataset.target_lufs. -20 LUFS leaves headroom and matches voice training defaults. (default -20; minimum -30; maximum -10)

Sample rate - dataset.sample_rate. 24000 Hz is required by the IndexTTS training pipeline. (default 24000; minimum 8000; maximum 48000)

Minimum file alignment coverage - dataset.min_file_alignment_coverage. Below 0.60, sentence alignment falls back or rejects unreliable timing. (default 0.6; minimum 0; maximum 1)

Minimum segment alignment coverage - dataset.min_segment_alignment_coverage. Drops individual caption segments with weak word alignment. (default 0.7; minimum 0; maximum 1)

Minimum words / second - dataset.min_words_per_second. Drops unusually sparse transcript/audio matches. (default 1; minimum 0.1; maximum 5)

Maximum words / second - dataset.max_words_per_second. Drops implausibly dense or misaligned speech. (default 5.5; minimum 1; maximum 12)

Minimum peak dBFS - dataset.min_peak_dbfs. Drops audio too quiet to train reliably. (default -35; minimum -80; maximum 0)

Maximum clipping ratio - dataset.max_clipping_ratio. 0.001 allows at most 0.1% clipped samples. (default 0.001; minimum 0; maximum 1)

Clipping threshold - dataset.clipping_threshold. Absolute normalized sample level counted as clipping. (default 0.999; minimum 0.5; maximum 1)

Maximum silence ratio - dataset.max_silence_ratio. Optional; blank disables whole-segment silence-ratio filtering. (default blank; minimum 0; maximum 1)

Silence threshold dBFS - dataset.silence_threshold_dbfs. Frames below this level count as silence. (default -40; minimum -80; maximum -10)

Silence frame (ms) - dataset.silence_frame_ms. 20 ms gives stable silence estimates for speech. (default 20; minimum 5; maximum 200)

Reference candidates - dataset.export_reference_candidates. Exports the cleanest segments for training samples and LoRA / DoRA use. (default 5; minimum 0; maximum 20)

Overwrite dataset - dataset.overwrite. Replaces an existing dataset directory with the same name. (default false)

Maximum segments - dataset.max_segments. 0 processes all segments; use a small value for smoke tests. (default 0; minimum 0; maximum 10000000)

Preparation seed - dataset.seed. Controls deterministic candidate ranking and randomized operations. (default 0; minimum 0; maximum 4294967295)

LoRA / DoRA Training - 87 settings

Dataset - training.dataset_dir. Prepared manifest dataset used for cached-feature training. (default "datasets/secourses_demo")

LoRA / DoRA name - training.name. Safe output folder and final safetensors basename. (default "voice_adapter")

LoRA / DoRA type - training.adapter_type. DoRA is the quality default; LoRA uses slightly less compute. (default "dora"; choices "lora", "dora")

Rank - training.rank. 128 with alpha 129 learned the voice fastest in measured runs; rank 32 reached the same floor more slowly. (default 128; minimum 1; maximum 256)

Alpha - training.alpha. 129 is the measured companion scale for the recommended rank 128. (default 129; minimum 1; maximum 1024)

Dropout - training.dropout. 0.05 remains the quality default; stronger measured regularization gave no benefit. (default 0.05; minimum 0; maximum 0.5)

Target attention - training.target_attention. Adapts GPT attention projections; recommended. (default true)

Target MLP - training.target_mlp. Adapts GPT feed-forward projections; recommended for voice fidelity. (default true)

Train speaker projection - training.train_spk_proj. Fully trains the small speaker projection module. (default true)

Train emotion layers - training.train_emo_layers. Advanced: trains small emotion modules in addition to LoRA / DoRA layers. (default false)

Train mel embedding head - training.train_mel_embed_head. Advanced: trains the mel token embedding/head modules. (default false)

Learning rate - training.learning_rate. 4e-5 is the robust batch-1 default: it reached 5.061 held-out loss, while 8e-5 overfit within two epochs. (default 0.00004; minimum 1e-8; maximum 1)

Optimizer - training.optimizer. AdamW is portable; fused AdamW is faster on supported CUDA builds. (default "adamw"; choices "adamw", "adamw_fused", "prodigy")

Scheduler - training.lr_scheduler. Cosine decay is recommended for multi-epoch voice adaptation. (default "cosine"; choices "cosine", "linear", "constant", "constant_with_warmup")

Warmup steps - training.warmup_steps. 200 steps is the measured batch-1 default, easing training into the 4e-5 learning rate. (default 200; minimum 0; maximum 1000000)

Weight decay - training.weight_decay. 0.01 is a mild regularizer; 0.05 showed no measured benefit. (default 0.01; minimum 0; maximum 1)

Adam betas - training.betas. Two comma-separated momentum coefficients; 0.9, 0.99 is recommended. (default "0.9, 0.99")

Adam epsilon - training.eps. Numerical stability term for Adam-family optimizers. (default 1e-8; minimum 1e-12; maximum 0.1)

Epochs - training.epochs. 10 epochs is the measured batch-1 default; 20 epochs overfit after the held-out optimum around epoch 6. (default 10; minimum 1; maximum 10000)

Maximum steps - training.max_steps. 0 derives steps from epochs; set 5 for a quick smoke run. (default 0; minimum 0; maximum 100000000)

Batch size - training.batch_size. 1 is the measured quality default: every training clip becomes an optimizer update each epoch. (default 1; minimum 1; maximum 128)

Gradient accumulation - training.grad_accumulation. 1 is the measured default; accumulation 2 or 4 removed updates and performed worse at the same learning rate. (default 1; minimum 1; maximum 128)

Gradient clip - training.max_grad_norm. 1.0 limits unstable gradient spikes; 0 disables clipping. (default 1; minimum 0; maximum 100)

Label smoothing - training.label_smoothing. 0 is recommended; increase only for overconfident large datasets. (default 0; minimum 0; maximum 0.5)

Mel loss weight - training.mel_loss_weight. Primary autoregressive acoustic-token loss weight. (default 1; minimum 0; maximum 100)

Text loss weight - training.text_loss_weight. Auxiliary text modeling loss weight. (default 0.1; minimum 0; maximum 100)

Speaker reference mode - training.speaker_ref_mode. self uses the target clip, other uses a deterministic different same-speaker clip, and mixed alternates between them; other is the measured quality default. (default "other"; choices "self", "other", "mixed")

Emotion reference mode - training.emo_ref_mode. self uses the target emotion, other uses another same-speaker clip, mixed alternates, and follow_speaker reuses the speaker-reference clip; follow_speaker is the measured inference-like default. (default "follow_speaker"; choices "self", "other", "mixed", "follow_speaker")

Maximum codes - training.max_codes. Cached samples longer than this semantic-code limit are rejected. (default 1500; minimum 1; maximum 100000)

Maximum text tokens - training.max_text_tokens. Cached text length safety limit. (default 600; minimum 1; maximum 100000)

Validation fraction - training.val_fraction. 5% provides useful validation without sacrificing much training data. (default 0.05; minimum 0; maximum 0.5)

Validate every steps - training.val_every_steps. 0 disables step validation; epoch validation still runs when a split exists. (default 50; minimum 0; maximum 1000000)

Maximum validation batches - training.val_max_batches. Caps validation time on large datasets. (default 20; minimum 1; maximum 1000000)

Validation reference - training.val_reference_mode. self validates each target with itself, while other uses a different same-speaker clip for both vectors; other is inference-like and measured more accurately. (default "other"; choices "self", "other")

Early-stop patience - training.early_stop_patience. 0 disables early stopping; otherwise stop after this many validations without a meaningful improvement. (default 0; minimum 0; maximum 1000000)

Early-stop minimum improvement - training.early_stop_min_delta. Validation loss must fall by at least this amount to reset patience. (default 0; minimum 0; maximum 1000000)

Base variant - training.base_variant. BF16 is the quality default; INT8 ConvRot reduces frozen base weight memory. (default "bf16"; choices "bf16", "int8_convrot")

Base dtype - training.base_dtype. Compute/storage dtype for the BF16 base variant. (default "bf16"; choices "bf16", "fp16", "fp32")

Mixed precision - training.mixed_precision. BF16 is recommended on modern GPUs and avoids FP16 overflow. (default "bf16"; choices "bf16", "fp16", "fp32")

Gradient checkpointing - training.gradient_checkpointing. Recommended; recomputes activations to save substantial VRAM. (default true)

Blocks to swap - training.blocks_to_swap. Streams this many frozen GPT blocks from CPU; requires gradient checkpointing. (default 0; minimum 0; maximum 24)

Swap ring size - training.swap_ring_size. 2 balances overlap and VRAM; 1 uses the least memory. (default 2; minimum 1; maximum 4)

Pinned swap memory - training.pin_swap_memory. Recommended for faster CPU-to-GPU transfers. (default true)

Output root - training.output_dir. LoRA / DoRA parent folder; relative paths resolve from the app directory. (default "loras")

Save every epochs - training.save_every_epochs. 1 keeps an epoch checkpoint; 0 disables epoch checkpoints. (default 1; minimum 0; maximum 100000)

Save every steps - training.save_every_steps. 0 disables step checkpoints. (default 0; minimum 0; maximum 10000000)

Keep last N - training.keep_last_n. 0 keeps every epoch checkpoint so measured checkpoint comparison can choose the best voice. (default 0; minimum 0; maximum 10000)

Save best - training.save_best. Keeps the checkpoint with the lowest validation loss. (default true)

LoRA / DoRA save dtype - training.save_dtype. BF16 halves LoRA / DoRA file size; FP32 preserves full update precision. (default "bf16"; choices "bf16", "fp32")

Save train state - training.save_train_state. Saves optimizer, scheduler, scaler, RNG, and data position for exact resume from best, final, and interrupted checkpoints. (default true)

Save train state with every epoch checkpoint - training.epoch_train_state. Only needed to Continue run from a specific epoch; costs ~4x disk per checkpoint. (default false)

Resume from - training.resume_from. Select a LoRA / DoRA checkpoint; rank, alpha, and type are inspected before launch.

Resume mode - training.resume_mode. Weights only starts a fresh schedule at step 0; Continue run restores train state when available. (default "weights_only"; choices "weights_only", "continue")

Analyze generalization automatically - training.auto_analyze. Reads the CPU-only training log after complete or stopped runs and recommends a checkpoint. (default true)

Evaluate checkpoints automatically - training.auto_evaluate_checkpoints. After training releases its model, measures saved checkpoints on validation and a small training subset. (default true)

Evaluate Base model (no LoRA / DoRA) - training.eval_include_base. Measures the reference-only baseline before checkpoints for an automatic comparison. (default true)

Evaluation training subset - training.eval_train_subset. Deterministic training items measured during automatic evaluation; 0 disables the training subset. (default 48; minimum 0; maximum 100000)

Evaluation strengths - training.eval_strengths. Comma-separated LoRA / DoRA strengths from 0 to 4 for automatic checkpoint evaluation. (default "1.0")

Evaluation timeout (s) - training.eval_timeout_s. Stops automatic checkpoint evaluation after this many seconds without failing the completed training run. (default 900; minimum 1; maximum 100000)

Generate training samples - training.sample_enabled. Renders a short sample at the configured epoch interval. (default true)

Sample every epochs - training.sample_every_epochs. 1 provides a sample after each completed epoch. (default 1; minimum 1; maximum 10000)

Sample runtime tier - training.sample_runtime_tier. Memory tier for the isolated sampling process. (default "auto"; choices "auto", "6", "8", "10", "12", "16", "24", "32")

Minimum free VRAM (GB) - training.sample_min_free_vram_gb. Skips sampling rather than risking training OOM below this free-memory threshold. (default 6; minimum 0; maximum 128)

Sample timeout (s) - training.sample_timeout_s. Kills a stuck sampling subprocess after this time. (default 300; minimum 1; maximum 100000)

Sample text - training.sample_text. Short representative phrase used to compare epochs. (default "This is a training progress sample for the adapted voice.")

Custom sample reference - training.sample_reference. Optional audio path; blank uses the dataset's best reference candidate automatically.

Sample language - training.sample_language. Mirrors Voice Generation for per-epoch samples; auto uses the prepared dataset language. (default "auto"; choices "auto", "ZH", "EN", "JA", "AR", "ES")

Sample seed - training.sample_seed. Mirrors Voice Generation for per-epoch samples; -1 chooses one seed at training start and reuses it across epochs. (default -1; minimum -1; maximum 4294967295)

Sample beams - training.sample_num_beams. Mirrors Voice Generation beam search for every per-epoch sample. (default 3; minimum 1; maximum 10)

Sample temperature - training.sample_temperature. Mirrors Voice Generation temperature for every per-epoch sample. (default 0.8; minimum 0.1; maximum 2)

Sample top-p - training.sample_top_p. Mirrors Voice Generation nucleus sampling for every per-epoch sample. (default 0.8; minimum 0; maximum 1)

Sample top-k - training.sample_top_k. Mirrors Voice Generation token filtering for per-epoch samples; 0 disables it. (default 30; minimum 0; maximum 100)

Sample repetition penalty - training.sample_repetition_penalty. Mirrors Voice Generation repetition control for every per-epoch sample. (default 10; minimum 1; maximum 20)

Sample emotion weight - training.sample_emo_alpha. Mirrors Voice Generation emotion weight for every per-epoch sample. (default 0.65; minimum 0; maximum 1)

Sample diffusion steps - training.sample_diffusion_steps. Mirrors Voice Generation diffusion quality for every per-epoch sample. (default 25; minimum 2; maximum 100)

Sample CFG rate - training.sample_inference_cfg_rate. Mirrors Voice Generation conditioning strength for every per-epoch sample. (default 0.7; minimum 0; maximum 2)

Sample maximum text tokens - training.sample_max_text_tokens. Mirrors Voice Generation segment length for every per-epoch sample. (default 60; minimum 20; maximum 300)

Sample length penalty - training.sample_length_penalty. Mirrors Voice Generation beam length control for every per-epoch sample. (default 0; minimum -2; maximum 2)

Sample maximum mel tokens - training.sample_max_mel_tokens. Mirrors Voice Generation output limit for every per-epoch sample. (default 1500; minimum 1; maximum 1815)

Sample speaking rate - training.sample_speaking_rate. 1.0 is the model's natural pace; below 1.0 speaks slower and above 1.0 faster. (default 1; minimum 0.5; maximum 1.5)

Seed - training.seed. Controls split, sampler, initialization, and training randomness. (default 42; minimum -2147483648; maximum 4294967295)

Data workers - training.num_workers. 2 is a safe Windows/Linux default; use 0 to debug worker issues. (default 2; minimum 0; maximum 64)

Log every steps - training.log_every_steps. 1 gives fully live charts; increase slightly for very fast runs. (default 1; minimum 1; maximum 100000)

Training device - training.device. CUDA device used by the training worker. (default "cuda:0")

Attention backend - training.attention_backend. SDPA is the compatible default. (default "sdpa"; choices "sdpa", "eager", "flash_attention_2")

Model directory - training.model_dir. Base IndexTTS 2.5 model directory. (default "models")

Model config - training.model_config. IndexTTS 2.5 YAML configuration path. (default "models/config.yaml")

Checkpoint Grid - 20 settings

LoRA / DoRA folder - grid.adapter_dir. Choose one training run to analyze and compare.

Evaluation references - grid.eval_reference_mode. Same as training validation inherits that run's validation mode; other uses a different same-speaker clip like inference. (choices "", "self", "other")

Training subset - grid.eval_train_subset. Deterministic training items measured beside validation; 0 disables the training subset. (default 48; minimum 0; maximum 100000)

Include base model - grid.eval_include_base. Adds Base model (no LoRA / DoRA), the reference-only comparison baseline. (default true)

Checkpoints - grid.checkpoints. Base model (no LoRA / DoRA), the recommended checkpoint, and the final checkpoint are selected first. (default [])

Strengths - grid.strengths. Comma-separated LoRA / DoRA strengths from 0 to 4; Base model (no LoRA / DoRA) is generated once without a strength. (default "1.0")

Texts (one per line) - grid.texts. Each non-empty line is generated for every selected row and reference. (default "This is a training progress sample for the adapted voice.\nA calm, natural voice should remain clear on a sentence it never heard during training.")

Reference audio paths (one per line) - grid.references. Every checkpoint uses these same references in the same order.

Seed - grid.seed. -1 draws one random seed at grid start and then fixes it across cells. (default -1; minimum -1; maximum 4294967295)

Language - grid.language. Use the language of every sentence in this grid. (default "EN"; choices "ZH", "EN", "JA", "AR", "ES")

Temperature - grid.temperature. 0.8 matches Voice Generation defaults. (default 0.8; minimum 0.1; maximum 2)

Top-p - grid.top_p. Nucleus sampling threshold shared by every cell. (default 0.8; minimum 0; maximum 1)

Top-k - grid.top_k. Token candidate cutoff; 0 disables it. (default 30; minimum 0; maximum 100)

Beams - grid.num_beams. Beam count strongly affects quality as well as generation time and VRAM. (default 3; minimum 1; maximum 10)

Repetition penalty - grid.repetition_penalty. Keeps semantic-token loops under control. (default 10; minimum 1; maximum 20)

Diffusion steps - grid.diffusion_steps. 25 is the quality default. (default 25; minimum 2; maximum 100)

CFG rate - grid.inference_cfg_rate. Diffusion conditioning strength. (default 0.7; minimum 0; maximum 2)

Max text tokens per segment - grid.max_text_tokens_per_segment. Use the same segment size for a fair comparison. (default 60; minimum 20; maximum 300)

Emotion weight (alpha) - grid.emotion_weight. Shared emotion-conditioning blend for every cell. (default 0.65; minimum 0; maximum 1)

Speaking rate - grid.speaking_rate. 1.0 is the model's natural pace; below 1.0 speaks slower, above 1.0 faster. A trained LoRA / DoRA can carry a calibrated value that matches the speaker's real pace. (default 1; minimum 0.5; maximum 1.5)

Models & Performance - 26 settings

LoRA / DoRA - runtime.lora_path. Select a trained LoRA / DoRA, or None for Base model (no LoRA / DoRA), which clones from the reference only.

LoRA / DoRA strength - runtime.lora_strength. 1.0 is the trained strength; lower is subtler and higher is stronger. (default 1; minimum 0; maximum 2)

Merge LoRA / DoRA into base weights for speed (BF16 only) - runtime.lora_merge_into_base. Temporarily folds the selected LoRA / DoRA into floating GPT weights and restores them before switching. (default false)

Device - runtime.device. Auto selects the first available accelerator; choose CPU only for diagnostics. (default "auto"; choices "cuda:0", "cuda:1", "auto", "cpu")

VRAM tier - runtime.vram_tier. Auto detects physical VRAM; selecting a named tier fills conservative runtime settings. (default "auto"; choices "auto", "6", "8", "10", "12", "16", "24", "32", "custom")

GPT model variant - runtime.model_variant. BF16 gives the official quality path; INT8 ConvRot reduces GPT weight memory. (default "bf16"; choices "bf16", "int8_convrot")

GPT dtype - runtime.gpt_dtype. BF16 is recommended on modern NVIDIA GPUs; FP32 is the CPU-compatible fallback. (default "bf16"; choices "bf16", "fp16", "fp32")

Attention backend - runtime.attention_backend. SDPA is the compatible default; FlashAttention 2 requires its optional package. (default "sdpa"; choices "sdpa", "flash_attention_2", "eager")

Use acceleration engine - runtime.use_accel. Enables the optional CUDA-graph/flash-attention path; use beams=1. (default false)

Enable emotion-text model - runtime.use_qwen_emo. Required for Emotion text mode; on-demand residency keeps startup lazy. (default true)

Compile s2mel - runtime.torch_compile_s2mel. Uses torch.compile for repeated workloads; first generation takes longer. (default false)

BigVGAN CUDA kernel - runtime.use_cuda_kernel_bigvgan. Uses the optional fused activation kernel when available. (default false)

CFM estimator BF16 autocast - runtime.s2mel_estimator_autocast. Reduces activation VRAM; useful at 6 GB and sometimes slightly changes output. (default false)

Use DeepSpeed loader - runtime.use_deepspeed. Optional legacy loader; leave off unless DeepSpeed is installed. (default false)

GPT blocks to swap - runtime.blocks_to_swap. 0 keeps all blocks resident; -1 lets runtime fit automatically; up to 24 streams from CPU. (default 0; minimum -1; maximum 24)

Swap ring size - runtime.swap_ring_size. 2 overlaps transfer and compute; 1 uses least VRAM. (default 2; minimum 1; maximum 4)

Pinned swap memory - runtime.pin_swap_memory. Recommended for faster CPU-to-GPU block transfers. (default true)

Runtime CFM cache length - runtime.cfm_cache_length. Upper cache reservation used when generation does not request a larger value. (default 8192; minimum 1024; maximum 32768)

VRAM reserve (GB) - runtime.vram_reserve_gb. 2 GB is recommended to absorb allocator and generation peaks. (default 2; minimum 0; maximum 12)

Section batch hint - runtime.max_section_batch_size_hint. Advisory maximum shown to generation controls for this runtime. (default 8; minimum 1; maximum 64)

Semantic Model - runtime.aux_residency.semantic_model. Residency policy for this auxiliary model. (default "gpu"; choices "gpu", "on_demand", "cpu")

Qwen Emo - runtime.aux_residency.qwen_emo. Residency policy for this auxiliary model. (default "on_demand"; choices "gpu", "on_demand", "cpu")

Campplus - runtime.aux_residency.campplus. Residency policy for this auxiliary model. (default "gpu"; choices "gpu", "on_demand", "cpu")

Semantic Codec - runtime.aux_residency.semantic_codec. Residency policy for this synthesis model. (default "gpu"; choices "gpu", "on_demand")

S2Mel - runtime.aux_residency.s2mel. Residency policy for this synthesis model. (default "gpu"; choices "gpu", "on_demand")

Bigvgan - runtime.aux_residency.bigvgan. Residency policy for this synthesis model. (default "gpu"; choices "gpu", "on_demand")

18. Final Checklist

One short base-model generation succeeds before advanced tuning.
The selected language matches the script and the reference is clean.
The active adapter, strength, auto-reference, and speaking rate are intentional.
The section preview, caption timing, pauses, and target-duration semantics match the use case.
The selected VRAM tier has passed the benchmark and section batch size stays within its hint.
Dataset segments were listened to before caching and training.
A checkpoint was chosen by validation plus fixed-input listening, not file recency alone.
The working state is saved as a named user preset and the output folder contains reproducibility metadata.

You now have a complete path from an untrained reference clip to repeatable production output, plus the tools to build and validate a dedicated voice adapter when reference-only cloning is not enough.

SwarmUI Latest Updates

Furkan Gözükara — Thu, 27 Aug 2026 22:54:57 +0000

By SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts | Original Patreon post

Get the SwarmUI Installer and Model Downloader app and the presets from here : https://www.patreon.com/SECourses/posts/swarmui-auto-and-114517862

28 August 2026 Update V172

Reference tokens are more robust now.
- Variants such as @IMAGE1, @ image # 2, and <picture1> are recognized and normalized automatically.
Prompt-label handling and makes token estimates more accurate for multi-frame continuation and Image-to-Video workflows now.
Better Video Continuation
- Init Video Continuation 1.3.1 now supports 1, 5, 22, 39, or 56 context frames.
- One-frame mode preserves the previous behavior. Multi-frame mode uses MiniMax H3's native clip conditioning for smoother transitions.
- Replayed context frames and matching audio are removed automatically during merging, preventing duplicated footage or sound.
- Source audio is preserved and generated audio is appended at the continuation boundary.
Start Frame / Init Video now accepts an image or video. Uploaded videos can be continued and optionally merged with the newly generated segment.
Continuation no longer consumes a Ref2VA reference-image slot.
MiniMax H3 Fixes
- Fixed text-only Audio Only workflows failing to recognize SwarmUI's H3 latent canvas.
- Fixed pasted multiline wildcard text being collapsed.
WD14 Tagger Hardening
- All 19 supported model repositories are now allowlisted and pinned to reviewed, immutable Hugging Face revisions.
- Downloaded Taggerine inference code must pass SHA-256 verification before execution.
- Added stricter validation for model IDs, thresholds, image types and sizes, model directories, and temporary output paths.
- Runtime package installation was removed. Missing dependencies now produce clear installation instructions instead of modifying the ComfyUI environment during generation.
- Added dedicated security-boundary tests.
- Install from this repo manually https://github.com/FurkanGozukara/SwarmUI-WD14Tagger
Automatic Installers / Updaters Reliability
- Aggressive Git updates now clean untracked leftovers after repository resets, improving recovery from blocked or inconsistent node updates.
Get latest zip file extract and overwrite all files
- Import latest presets
- Run Windows_Install_SwarmUI.bat to update latest
Get latest ComfyUI and update it to latest as well
- https://www.patreon.com/SECourses/posts/comfyui-auto-2-105023709
More latent frames will increase vram usage a lot so be careful
This feature added to ComfyUI presets as well read changelogs on ComfyUI

22 August 2026 Update V170

This is a very big update : LTX 2.5 video generation presets + LTX 2.5 Video Core Bundle, new FLUX 2 Klein INT8 / INT4 ConvRot HQ models, MiniMax H3 Init Audio (make the video follow any soundtrack with lipsync), live prompt token meter, Face Inpainting upgrades, a big model downloader UI upgrade and a huge installer / updater overhaul so please read all
LTX 2.5 video generation presets finally added : Amazing_SwarmUI_Presets_v70.json brings 6 new presets : LTX25 Text To Video - Core 2-Stage 8+3, LTX25 Image To Video - Core 2-Stage 8+3, LTX25 First Last Frame - Core 2-Stage 8+3, LTX25 Audio Reference To Video - Core 2-Stage 8+3, LTX25 Text To Video - Dev HQ 30+6 and LTX25 Image To Video - Dev HQ 30+6 (all 260822)
Audio Reference To Video : add a Prompt Audio file and describe the matching visuals - the native LTX 2.5 audio branch makes the video follow your audio (lipsync, timing) - Image To Video and First Last Frame presets : add an Init Image (and a Video End Frame) and describe only the motion, camera and sound instead of re-describing the still

All 6 presets are two-stage : the first stage generates at 960x544, then SwarmUI's Refine / Upscale applies the official LTX 2.5 latent spatial upscaler x2 and refines to model-aligned 1080p (1920x1088) with synchronized audio - 97 frames at 24 FPS (about 4 seconds) and H264 MP4 output by default
Core presets use my LTX 2.5 Distilled INT8 ConvRot Premium model with 8 + 3 steps at CFG 1 and Euler Ancestral so they are fast - Dev HQ presets use my LTX 2.5 Dev INT8 ConvRot Premium model with 30 + 6 steps at CFG 3 and Euler for maximum prompt adherence and detail
Everything is exposed as normal SwarmUI parameters so you can change resolution, duration, steps and the refiner settings as you like - increase Text2Video Frames for longer clips
Import Amazing_SwarmUI_Presets_v70.json or run Windows_Preset_Delete_Import.bat to get them - the Which_Bundles_Downloads_Which_Preset_Models.html report is regenerated for the new presets as well

LTX 2.5 Video Core Bundle added to the model downloader : 1 click downloads everything the LTX 2.5 presets need - 10 files, 64.64 GB : Dev and Distilled INT8 ConvRot Premium transformers, Gemma 4 12B INT8 ConvRot v2 text encoder (SwarmUI's current default), Gemma 4 E2B INT8 ConvRot prompt enhancer, the convolutional and the diffusion video VAE, the audio VAE, the latent spatial upscaler x2, the pixel spatial upscaler x2 IC-LoRA and the motion track control IC-LoRA
This one bundle covers every model referenced by all of our LTX 2.5 presets in both SwarmUI and ComfyUI, so you do not need to hunt for a second bundle to make any of them run
The text encoder and both VAEs are saved with the exact LTX-2 sub folder and file names SwarmUI's own automatic downloader expects, so SwarmUI never downloads a second copy - in ComfyUI mode they go to models/text_encoders/LTX-2 and models/vae/LTX-2
No more gated downloads for LTX 2.5 : I have mirrored the remaining official Lightricks files into my own non-gated repository, so nothing in the LTX 2.5 list asks you to log in or accept a license anymore - this covers the official Dev and Distilled Comfy INT8 ConvRot transformers, the Distilled NVFP4 transformer (for RTX 5000 Blackwell), the pixel spatial upscaler x2 IC-LoRA and the latent spatial upscaler
These official quantizations stay optional alternatives - my own Premium INT8 ConvRot conversions remain the default of the bundles and presets because they are higher quality, and the Gemma 4 E2B INT8 ConvRot prompt enhancer is also listed separately

Model downloader UI upgrade : every model list and every bundle Includes list is now grouped by the folder the files are saved into : Main Models, Text Encoders, VAEs, LoRAs, ControlNets, Upscalers, Vision, Detection, LLM - biggest files first, and every group header shows the file count and total GB so you instantly see what goes where

New Active Transfers panel with a real progress bar per file (percent, downloaded / total, speed, ETA) above the Activity Log - the log is chronological now with a Follow newest log entry checkbox, it refreshes twice per second, and multiple open browser tabs no longer steal each other's updates
The URL Downloader tab now shows its progress in the same panel instead of looking frozen, and every download reports [STARTING] immediately instead of showing nothing while the file metadata is fetched
Console progress line reordered : percent, downloaded / total size, speed and ETA come first and the long file name goes last, so a narrow CMD window never cuts off the numbers - and when the output is not a real terminal (RunPod / Massed Compute logs) progress is still printed once per second instead of being silent

New LTX2.5_Enchance_Prompt_Feed_For_LLMs.txt inside the Prompt_Generate_LTX_MiniMax_And_Presets_How_To_Use folder : an official-aligned LTX 2.5 prompting guide for LLMs - give it to ChatGPT / Claude / Gemini, tell it what you want, and it writes properly formatted LTX 2.5 prompts (text / image / audio to video, multi-shot and IC-LoRA conventions are all covered)
MiniMaxH3References extension 1.13.2 : videos and audio files attached with SwarmUI's own native prompt uploader are now recognized as MiniMax H3 references too - same cards, same @video1 / @audio1 tokens with autocomplete, same Ref2VA workflow
- All extensions and custom nodes automatically installed by our installers
FLUX 2 Klein INT8 and INT4 ConvRot HQ models : I have compiled high quality INT8 ConvRot and INT4 ConvRot versions of FLUX 2 Klein myself - Base 9B, Distilled 9B and Distilled 9B kv - 6 new models added to the model downloader app
FLUX 2 Klein Core Bundle and Complete Image Generation and Editing Bundle now download the INT8 ConvRot HQ versions - almost same quality as BF16 but massively faster, same as our Krea 2 and LTX 2.5 ConvRot models

With Amazing_SwarmUI_Presets_v70.json both FLUX 2 Klein Base and FLUX 2 Klein Distilled 8 Steps presets now use the INT8 ConvRot HQ models
MiniMax H3 FL2V Turbo 4-Step 768p speed LoRA updated to v1.1 : the model downloader now downloads minimax_h3_fl2v_turbo_4step_v1.1_768p_bf16.safetensors instead of the old v1.0 - this is the optional 16:9 aspect ratio speed up LoRA for 768p generations from LightX2V - MiniMax-H3 Core and Low VRAM bundles download the new version automatically

MiniMax H3 Init Audio : new Init Audio parameter group added - select or upload any audio file (or even a video file, its soundtrack is used) and MiniMax H3 generates the video to follow that soundtrack exactly : lipsync, action timing, ambience - and keeps it as the exact final audio track of the output
Works for Text To Video, Image To Video and References To Video presets and Init Audio Match Duration makes the video length automatically follow the audio length
Live token meter : a real time Tokens meter now appears above the prompt and shows estimated MiniMax H3 token usage (prompt + references + resolution + duration) against the 109k budget - you instantly see when your references or duration are too big before generating
Face Inpainting upgrades : new Face Inpaint Faces parameter selects which detected faces get refined : 1 = biggest face (default), 2 = second biggest, 1,3 or all - faces are ranked biggest to smallest and each selected face is refined in its own pass
New hallucination guard prevents pasting a wrong / neighbouring face over your subject and Face Inpaint Detector can no longer block generations

Installer / Updater overhaul : Windows_Install_SwarmUI.bat and Windows_Update_SwarmUI.bat are merged into a single Windows_Install_or_Update_SwarmUI.bat
Windows_Start_SwarmUI.bat is now ultra fast : it starts SwarmUI directly without updating or downloading anything and rebuilds only if the code actually changed
Update runs are much faster : unchanged extensions (Premium Extensions, Licon MSR, Foley) are skipped with a quick revision check instead of being re-downloaded on every run - FFmpeg and Cloudflared downloads are also skipped when already latest
Missing Git is now auto installed (winget on Windows, apt-get on Linux) instead of failing with cryptic WinError 2 errors
SwarmUI is now installed as a full git clone and older shallow installs are auto completed - fixes the scary Tag list empty?! warning of SwarmUI's own update check
.NET 10 SDK now installs into SwarmUI/.dotnet folder, matching SwarmUI's own launcher search order and persisting on RunPod / SimplePod network volumes - DOTNET_ROOT is exported so in-app Update and Restart now rebuilds with the correct SDK and works properly
Linux / cloud : Cloudflared is now installed system-wide from Cloudflare's official APT repository and launching prints the public trycloudflare.com URL so you can open SwarmUI from your own computer's browser on RunPod / SimplePod / Massed Compute - new --no-cloudflared and --host options added
RunPod, SimplePod and Massed Compute instruction files rewritten with exact HOW TO OPEN SWARMUI steps
Windows_Preset_Delete_Import.bat made more robust : works from any folder and auto picks the correct Python
To update : get latest zip file, extract and overwrite all, then run Windows_Install_or_Update_SwarmUI.bat and import Amazing_SwarmUI_Presets_v70.json
Also update your ComfyUI backend to latest V127 and get the 8 new LTX 2.5 ComfyUI presets : https://www.patreon.com/SECourses/posts/comfyui-auto-2-105023709

16 August 2026 Update V168

MiniMax H3 bundle updated to include newest 8-steps LoRA and also necessary Yolo Face model for new automatic ComfyUI Face Inpainting feature in presets for MiniMax H3
SwarmUI presets are now also using new Ref2V LoRA for references presets, you can switch to FL2V LoRA and see which one performs better so easy from LoRA tab and compare

Automatic video face inpainting implemented

Get latest zip file, extract and overwrite all, restart PC, run Windows_Update_SwarmUI.bat to update
Also import latest Amazing_SwarmUI_Presets json file
You can see impact of Face Inpainting here and now even better working : https://www.reddit.com/r/comfyui/comments/1vpl5bb/secourses_comfyui_installer_and_ready_presets_now/

14 August 2026 Update V166

We have added LTX 2.5 video upscaler and MiniMax Music 3 Text to Music presets into ComfyUI : https://www.patreon.com/SECourses/posts/comfyui-auto-2-105023709
- LTX 2.5 video upscaler preset is able to upscale any video, any resolution, any aspect ratio into 2x
- It is a very serious generative upscale so adds lots of details
- Therefore, try to upscale at 1 chunk to have better consistency
- Read preset information carefully after loading into ComfyUI
- Our SECourses Premium Upscaler Pro app also fully supports LTX 2.5 now : https://www.patreon.com/SECourses/posts/secourses-pro-150202809
  - It is standalone does not require ComfyUI and it auto downloads models when you first time use
To make these presets work out of the box, 2 new ComfyUI core bundles added to the downloader app as below
LTX 2.5 video generation presets and bundles will be added soon hopefully but its quality lower than MiniMax H3
Moreover, I have converted the LTX 2.5 Int8 ConvRot myself higher quality and more accurate than officially published ones
- Someone detected one of the model is corrupt and broken - officially released ConvRot Int8 variant
- I converted them from BF16 versions after big research and experimentation
- So our model downloader app will download better quality Int8 ConvRot of LTX 2.5 models
Our Musubi Trainer convert to quant tab also now supporting LTX 2.5 Int8 ConvRot compile : https://www.patreon.com/SECourses/posts/secourses-musubi-137551634

MiniMax Music 3 is reportedly lower quality than ACESTEP XL 1.5
- Our very advanced ACESTEP XL 1.5 Premium app below and it has full LoRA and DoRA training support
  - App : https://www.patreon.com/SECourses/posts/ace-step-1-5-xl-157675060
- MiniMax Music 3 also missing necessary resources to be trainable
- MiniMax Music 3 prompt templates shared here : https://github.com/MiniMax-AI/MiniMax-Music3/tree/main/skills/music-caption-rewriter/templates
To get updated model downloader app, as usual, get latest zip file, extract and overwrite all

12 August 2026 Update V165

The most professional speed up LoRA maker LightX2V just published new speed LoRAs for the MiniMax H3
- https://github.com/ModelTC/Minimax-H3-Turbo

Our MiniMax H3 core bundles now auto downloads both FL2VA Turbo 8-step v1.0 and FL2VA Turbo 4-step v1.0 768p
- Older v0.1 removed from downloads and presets so you can delete it from LoRA downloads if you did download before
- All presets now by default uses FL2VA Turbo 8-step v1.0 but you can switch it to FL2VA Turbo 4-step v1.0 768p
  - If you switch to FL2VA Turbo 4-step v1.0 768p you may be needed to overwrite Video shift value and set to 6, SwarmUI default sets to 12
- Sadly LoRAs only made for non-reference model atm but works on both models until they publish specific new LoRA for reference model
Quick set megapixels feature implemented
- Based on selected aspect ratio, it will set resolution according to your set
Moreover, now it will display visualization of the output based on aspect ratio and resolution

To update get the latest zip file, extract and overwrite all, run Windows_Update_SwarmUI.bat and also use Windows_Preset_Delete_Import.bat to get latest updated presets
Moreover get latest ComfyUI backend zip file and also update it through installer bat file : https://www.patreon.com/SECourses/posts/comfyui-auto-2-105023709

10 August 2026 Update V162

ComfyUI infinite video with MiniMax H3 tutorial published : https://youtu.be/1580ZDX-60Q

Hopefully a tutorial for SwarmUI coming soon

Completely Rebuilt Download Engine
- Exact byte-range resume after cancellation, connection loss, or application restart.
- Previously downloaded ranges are preserved instead of downloaded again.
- One sparse staging file replaces multiple temporary chunks and the expensive final merge.
- Significantly lower temporary disk-space requirements.
- Models are atomically installed only after size and SHA-256 verification.
- Corrupt or incomplete downloads can never replace a finished model.
Faster Hugging Face Downloads
- Native Hugging Face Xet is now recommended and enabled automatically when available.
- Xet automatically scales between 8 and 32 streams according to available system RAM.
- Improved recovery from stalled transfers, rate limits, server errors, and expired signed URLs.
- Removed the previous forced approximately 84.5 GiB Xet cache.
- New downloads default to no chunk cache, with an optional bounded reuse mode.
Improved Download Controls
- Direct URL downloads now appear in the active-download counter.
- Cancel and Cancel All now work with direct URL downloads.
- Completed ranges remain available after cancellation.
- Prevented overlapping standalone URL downloads.
- Small and large direct downloads now use the same atomic download system.
- Improved URL validation and rejection of HTML pages masquerading as model files.
- Safer Hugging Face and CivitAI domain detection.
Catalog And Interface Improvements
- Replaced the unavailable LTX 2.3 x2 spatial upscaler v1.0 with the official v1.1 long-video hotfix : ltx-2.3-spatial-upscaler-x2-1.1.safetensors
- Removed duplicate physical models from search and LoRA displays while keeping aliases searchable.
- Improved automatic folder detection for filenames containing hyphens and underscores.
- Improved mobile and narrow-screen layout.
Updater now detects and repairs incomplete FoleyExtension installations, including optional-image components.
MiniMax H3 Improvements
- Expanded the MiniMax H3 prompt-enhancement guide to v2.3.
- Added stable reference-roster numbering rules for multi-scene and folder-batch prompts.
- Added clearer image, video, audio, and paired-soundtrack reference rules.
- Corrected first-and-last-frame preset instructions to use Image To Video > Video End Image.
For updating please also update ComfyUI to latest as well and also this one
Use Windows_Update_SwarmUI.bat and import latest presets file

9 August 2026 Update V160

Model downloader app updated and Int4 ConvRot MiniMax H3 models bundle added as Low VRAM
- This bundle is great for 12 GB and below GPUs
- At the presets both ComfyUI and SwarmUI, just replace Int8 Model with below Int4 variants
- This format Int4 ConvRot models right out of the box working with our ComfyUI (use our installers) backend and installers for SwarmUI

We are using official code of Lightricks not Kijai implementation therefore our LoRA works better
However, Kijai had a VRAM optimization and i just implemented it
You will see it like below in presets, MAX savings reduces VRAM more than 40% but may reduce quality, exact saves 15%+ but same quality
Make sure to use our ComfyUI installer and update it for all features to run via our bat file
Remember MiniMax H3 specific features will only appear when you have selected MiniMax H3 architecture model in your model selection

If you set reference image size to max like this, it improves quality and accuracy but may use more memory, default is match

The importance of Max option is very significant

Video clip references are processed differently and below explains how they processed when used as a reference

8 August 2026 Update v159

This is a very major upgrade with so many new amazing stuff so please read carefully
Famous Lightricks released 4 steps LoRA for MiniMax H3
- https://github.com/ModelTC/Minimax-H3-Turbo
- https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main
- I have fully and accurately implemented it into our installers
  - Currently not working with SwarmUI or ComfyUI as default but works with our installers due to myself developed premium nodes and extensions
  - Make sure to install update your ComfyUI and SwarmUI users must also use our ComfyUI backend
- This LoRA added into our model downloader MiniMax H3 core bundle please download

I have made new presets that uses this LoRA and does 8 steps so now you can use them
- It can go as low as 4 steps but I recommend 8 steps
New updated presets are as below

Audio Only preset is super fast and generates audio directly
- Audio generation can be used to see if your prompt and duration matching before generating video since it is like real 2x time speed
  - So 1 minute audio generation takes like 30 seconds on RTX 5090 therefore use audio generation preset to see if your prompt and duration matches - for speaking having videos
  - So quickly iterate, see if your audio accurate, then generate full video
Moreover, I recommend you to use 8 steps and 0.4 megapixel resolution to quickly generate your videos, verify they are accurate, then move full generation like 1344x768px and 20 steps high quality
- This works great
Add references and write prompt field improved so that now you can add videos and audios with trim as you wish as below (trimming is optional)
- For @video1's soundtrack, type <Audio 1>; @audio1 is the first standalone audio file
- According to the above rule, MiniMax_H3_Enchance_Prompt_Feed_For_LLMs.txt file improved and literally working amazing
- It is located inside Prompt_Generate_LTX_MiniMax_And_Presets_How_To_Use folder inside zip file
Trim interface supports both Audio and Video trimming

ComfyUI presets have amazing batch folder processing read changelogs : https://www.patreon.com/SECourses/posts/comfyui-auto-2-105023709
For SwarmUI, you can use Wildcards as a batch processing
- To process them with order not randomly, enable Display Advanced Options
- Then find Swarm Internal and change Wildcard Seed Behaviour to index
- Set regular seed 0 so it will start from first prompt in Wildcard and how many generations you make, it will make with order

For everything to work, make sure that you use our ComfyUI installer as a backend and it is updated : https://www.patreon.com/SECourses/posts/comfyui-auto-2-105023709
Then get latest SwarmUI zip file, extract and overwrite all, then run Windows_Update_SwarmUI.bat to update
- Both install and update installs necessary extensions automatically
To update presets, i recommend use Windows_Preset_Delete_Import.bat
- You can also manually import latest Amazing_SwarmUI_Presets json file and overwrite all

6 August 2026 V158 - Update 2

We have significantly improved our add references extension - auto installed with our installer
Now when you add or remove an attachment, it will not modify and alter your written prompt
Now you can drag and drop attachments to change their position like @image1 into @image3 etc
Now even if you have reference in the prompt but not in attachment, it will not throw error and at the back end it will remove @image1 etc and prompt will just run fine
Zip file is same just run windows update bat file
Also I published a speed up quality comparison video : https://www.reddit.com/r/comfyui/comments/1vh5wd8/i_have_implemented_solattn_crossstep_cache_from/

6 August 2026 V158

I have fully and accurately implemented https://nvlabs.github.io/Sana/Sol-Engine/H3-OnDevice/ into our ComfyUI installation via our always installed and myself developed custom nodes
- Deploy MiniMax-H3 with Sol Engine achieving up to 4.52× acceleration on on-device GPUs
  - Their claimed 4.52x speed acceleration happens at 50 steps and default attention meanwhile I measured with 20 steps and against Sage Attention 2.8.3 latest (already speeds up like 20%+)
  - Moreover, I did set more conservative but almost same quality 0.08 thus it skips lesser steps
  - Use our ComfyUI backend installer and update it so this will work : https://www.patreon.com/SECourses/posts/105023709
Then updated our SwarmUI extensions to fully support it
Existing presets have a minor fix and they will work
When you select a MiniMax H3 architecture model, the parameters will appear like this - the backend has to be loaded so wait a little bit for it to appear

Default values quality wise conservative so almost same quality and brings 1.39x speed compared to Sage Attention 2 speeds at 1344x768 with 362 frames
It used only 20 GB vram for 1344x768 px at 362 frames - 15 seconds but it can use even lower depending on your GPU VRAM
The speed up is not enabled by default so check checkbox to enable it
The first run will be slow due to kernel compile but consquent runs will be faster and won't recompile
In logs first 4 steps are same, then step speed increases and then you start to see jumps in steps like below

For updating get latest zip file, overwrite all files, and then run Windows_Update_SwarmUI.bat - fresh install will also get this
- Make sure that you have updated your ComfyUI backend with our ComfyUI installer

5 August 2026 V157

We have officially moved enhanced SwarmUI with this update with the extensions I developed - all auto installed - please read carefully

MiniMax H3 is now fully supported with our Enhanced SwarmUI and ComfyUI installers and model downloaders
- Can generate up to 30 seconds and 2K native resolution just fine
In SwarmUI model downloader - download MiniMax H3 Core bundle
All of our presets both ComfyUI and SwarmUI, uses this bundle therefore downloading this will let you run presets right away
- Even if you have low VRAM GPU still this model working perfect
- Make sure to have at least 100 GB virtual RAM set
- Current GGUF releases are all almost same size as Int8 ConvRot therefore they are useless
- If Q4 proper one gets released I plan to add that into MiniMax-H3 Video Models section inside Video Generation Models tab

5 perfectly working new amazing MiniMax 3 presets with latest Amazing_SwarmUI_Presets_v62.json
- MiniMax H3 Beginning and Ending Image/Frame to Video - 260804
- MiniMax H3 Image to Video - 260804
- MiniMax H3 References to Video - 260804
- MiniMax H3 Text to Image with Optional References - 260804
- MiniMax H3 Text to Video - 260804
I recommend you to use Windows_Preset_Delete_Import.bat but you can always manually import too

Now recommended to add --disable-pinned-memory into your ComfyUI backend arguments until its bugs fixed to not have any OOM error
I have improved our special installers and extensions therefore you can use prompt input field to add 3 audio, 3 video and 9 image references to use with MiniMax H3 in SwarmUI easily
- They are referenced and named as @image1 @image2 @video1 @video2 @audio1 @audio2 and so on
The special upload will be allowed only when you select MiniMax H3 model
- It auto detects and enables based on selected model's architecture
Official SwarmUI still doesn't have this feature but we have with our installers
- We still always install official latest version SwarmUI but this comes from our own developed extension

New folder : Prompt_Generate_LTX_MiniMax_And_Presets_How_To_Use
- Our system uses references as @image1 @image2 @video1 @video2 @audio1 @audio2 and so on
  - Up to 9 images, 3 audios and 3 videos
  - <audio 1> means the audio of the first input <video 1>
- MiniMax_H3_Enchance_Prompt_Feed_For_LLMs.txt file is modified to instruct LLM to write prompts according to our system (as mentioned above) so you will get perfectly ready copy to paste prompts
New quick video extend feature
- Works with all of the init image to video models including MiniMax H3
- Apply Image to Video preset
- Then into init image field, instead of an image, input a video
- Then enable the Continue Init Video From Last Frame checkbox
- It will get last frame of your input video, generate image to video, and then finally merge videos and generate final output

Installers and update bat file made much more robust both for Windows and Linux
Get latest zip file, overwrite all, and run Windows_Update_SwarmUI.bat

19 July 2026 V154

This is a very big update with so many important new features, changes and improvements so please read all thank you

Seven New Presets

Make sure to import and overwrite Amazing_SwarmUI_Presets_v59.json or use Windows_Preset_Delete_Import.bat while SwarmUI running

Krea 2
- Krea 2 Raw/Base Slow - 260711 uses 52 steps, Euler, Simple, and CFG 4.5. It is the slower Raw/Base path intended mainly for fine-tuning, post-training, and LoRA work.
- Krea 2 Turbo Fast - 260711 uses 8 steps, Euler, Beta, and CFG 1 for fast final inference.
- Krea 2 Turbo Image Edit - 260716 adds denoise-based image editing with a default creativity value of 0.65. Add the source as Init Image and describe the requested change.

The new Krea 2 Core Bundle contains the Raw/Base INT8 HQ model, Turbo INT8 HQ model, Qwen3-VL 4B text encoder, and Qwen Image VAE. Its cached total is approximately 34.83 GiB.
The catalog also adds individually selectable Krea 2 Turbo Q8 GGUF and NVFP4 High Quality variants.

Automatically installed ComfyUI-QuantOps updated (With our ComfyUI Installer for SwarmUI Backend)
- Now it is only used if native ComfyUI is not supporting the loaded model
- I have compared the new famous Int8 ConvRot of Krea 2 and the speed difference is like 100%, I plan to update all models to Int8 ConvRot
- New updated ComfyUI-QuantOps supports Krea 2 GGUF as well

When you click and see full size of above image and analyze results you will see that:
- Int8 ConvRot is 96.2% similar to BF16 meanwhile GGUF Q8 is only 90.0% and FP8 Scaled is 82.2% and NVFP4 is 63.7%
  - Moreover, Int8 ConvRot generates the output in 3.05 seconds, making it 1.82× faster than BF16, which takes 5.56 seconds.
  - NVFP4 takes 3.8 seconds and is 1.46× faster than BF16, whereas GGUF Q8 takes 6.06 seconds and is approximately 8.3% slower than BF16.
  - So Int8 ConvRot generated with our Musubi Trainer app at high quality is almost 100% faster and almost same quality as BF16
  - High quality generation takes few hours on RTX 5090
Updated SwarmUI downloader has amazing quality Int8 ConvRot which reaches almost BF16 quality : https://www.patreon.com/SECourses/posts/swarmui-auto-and-114517862
- Krea 2 Core bundle downloads this model and uses it in SwarmUI preset
- ComfyUI preset also uses that model default
You can generate Int8 ConvRot models with our updated Musubi Trainer app : https://www.patreon.com/SECourses/posts/secourses-musubi-137551634

LTX 2.3 Foley Video-to-Audio

The new LTX23 Foley Video To Audio preset generates synchronized audio for a silent or muted video while keeping the source video unchanged.
Put the video in Init Image, then describe the visible sound source, action, material, and timing. The preset uses the LTX 2.3 Foley V2A LoRA with 30 steps, audio CFG 6, STG scale 1, modality scale 3, and a maximum of 169 frames at 24 FPS.
The Foley LoRA is approximately 0.21 GiB and is now part of the LTX 2.3 Video Core Bundle. The updated premium installer also installs the maintained FoleyExtension node package automatically.
- This model uses Dev version of LTX 2.3 not Turbo, therefore, now Dev version is included in the LTX 2.3 core bundle
For clean Foley generation, a useful prompt ending is: No speech is present. No music is present.
Download LTX 2.3 Core Bundle and in ComfyUI, install bundle 100

LTX 2.3 Licon MSR V2

The new LTX23 Licon MSR V2 Multi Subject Reference preset supports multiple subject, object, texture, or viewpoint references plus a required background image.
Add two to five images to Prompt Images in this exact order:
1. One to four subject, object, texture, or viewpoint references (so first 1-4 images are subjects)
2. The required background image last (last image is always background)
  1. Therefore, you need minimum 2 input images into the prompt field
Do not use Init Image for this workflow. Identify each numbered reference and its role in the positive prompt, then describe the new action, scene, camera, and lighting.
The preset uses the official eight-step distilled sigma schedule, Euler Ancestral sampling, a 65-frame internal reference sequence, and the LTX 2.3 Licon MSR V2 IC-LoRA. The new LoRA is approximately 0.61 GiB and is also included in the LTX 2.3 Video Core Bundle.
A dedicated guide is included with reference ordering, recommended settings, installation paths, and licensing details.
For this to work, we have coded custom extensions and all is automatically installer with SwarmUI installer or Updater files
Download LTX 2.3 Core Bundle and in ComfyUI, install Bundle 100 for ComfyUI backend

It is insanely powerful and fast I mean look below example

LTX 2.3 Presets Updates

LTX 2.0 model presets removed since LTX 2.3 is better in everyway
LTX 2.3 presets now using myself compiled Int8 ConvRot HQ models since they are literally 100% faster
- Model downloader LTX 2.3 core bundle is now downloading these new models

You can generate Int8 ConvRot models with our updated Musubi Trainer app : https://www.patreon.com/SECourses/posts/secourses-musubi-137551634

So Int8 ConvRot HQ is 100% faster than FP8 Quant Scaled and 50% faster than BF16 on RTX 5090
The quality is also excellent almost same as BF16

Phantom-Wan Character Reference T2V

Two Phantom-Wan presets are included:
- Phantom Wan 14B Character Reference T2V Fast
- Phantom Wan 14B Character Reference T2V Quality
Add character references to Prompt Images, not Init Image, then describe every subject, the scene, action, camera, and lighting. The reference images condition identity while generation begins from an empty video latent, so this is true reference-conditioned text-to-video rather than first-frame image-to-video.
One to four references are the officially recommended path. The integration can accept up to six for experimentation.
Download Phantom Wan 14B Character Reference T2V Bundle and in ComfyUI, install Bundle 100 for ComfyUI backend

Bundle 100 for ComfyUI Backend
- Inside ComfyUI zip file : Windows_Custom_Nodes_Bundles_Installer.bat

Unified Fast Robust Model Downloader Improvements

Run Windows_Start_Download_Models_App.bat to start Downloader
The previous hf_transfer checkbox has been replaced with two clearer controls:
- Custom Downloader Threads: 1 to 32, with 16 as the recommended default
- Hugging Face Xet Backend: optional and disabled by default
The custom thread setting applies to individual catalog downloads, bulk queues, bundles, search results, snapshots, and direct URL downloads. When Xet is enabled, the custom thread setting is ignored. If Xet fails, v150 falls back to the custom resumable backend unless the operation was cancelled.
The custom backend now has much stronger recovery behavior:
- Known-size partial downloads are preserved when cancelled.
- Parallel downloads save their range layout so later attempts can resume safely.
- Legacy v145 16-part downloads can be migrated and resumed.
- Every retry re-reads the saved byte count to prevent duplicate ranges and corruption.
- Six attempts use bounded exponential backoff.
- HTTP 429, 500, 502, 503, and 504 responses receive controlled retries.
- Separate connection and idle-read timeouts recover stalled transfers sooner.
- Progress output reports when it is waiting for the server and uses smoother speed estimates.
- Free space is checked before downloading and before merging parallel parts, with a safety reserve.
- Disk-full failures preserve useful partial data.
- Parallel parts are merged to a temporary file and atomically activated.
- Hugging Face main file URLs can use SHA verification and the optional Xet path.
SHA-256 verification and the verified-file cache remain in place. The goal is simple: fewer restarts from zero, clearer logs, and safer recovery on large model downloads.
Now you can set Hugging Face token for faster or private repo downloads
Now you can set HF Xet Download - even faster optional
Now you can set number of download threads
Now you can set Parallel File Downloads count
- This is extremely useful on high bandwith systems to download even faster than Hugging Face single file download limits like reaching 1 GB per second

Installation and Update Reliability

Windows install and update scripts no longer require one exact .NET SDK patch version. The new helper searches common system, user, registry, PATH, and environment locations for any usable stable .NET 10.x SDK while ignoring preview-only builds.
If installation is needed, it tries WinGet first and then falls back to the latest official Microsoft installer for x64, ARM64, or x86. The fallback download is checked against Microsoft's published release hash before it runs, and installer progress and log locations are shown instead of failing silently.
The premium-extension installer now validates the required integrations before changing the active setup. It installs or updates:
- FoleyExtension nodes plus the managed LTX 2.3 Foley SwarmUI integration
- Phantom character-reference helper nodes for the Comfy backend
- The managed Licon MSR SwarmUI integration
- The official Licon MSR Comfy node at the tested pinned revision
The new helper-node installers use staged directories, required-file checks, backup/rollback paths, bounded Git operations, and stale compiled-extension cleanup. Windows, RunPod, and Massed Compute update instructions now stop with an explicit error if this stage fails instead of continuing into a partial setup.

Updated Preset Model Report

A new file, Windows_Update_And_Open_Model_Report.bat, regenerates and opens the HTML report that shows which bundles and model files cover every SwarmUI preset.
The report generator now handles Krea 2, Phantom-Wan, Ideogram 4, negative-model parameters, saved-filename aliases, and LTX 2.3 Dev versus Distilled connector selection more accurately. The packaged report has been refreshed against Amazing_SwarmUI_Presets_v55.json.

How to Upgrade

Existing installation

Close SwarmUI and the model downloader.
Back up custom presets and any source changes you made inside the SwarmUI Git checkout. The updater resets tracked checkout files before pulling upstream changes.
Extract the latest zip file over your existing downloader package folder and allow release files to be replaced. Do not delete your SwarmUI or model directories.
Run Windows_Update_SwarmUI.bat.
Start the model downloader and download only the new bundles or individual files you need.
Import Amazing_SwarmUI_Presets_v55.json.

Windows_Preset_Delete_Import.bat can automate the preset replacement while SwarmUI is running. It first backs up the current presets, then deletes all active presets and imports the latest pack. Read its confirmation carefully if you maintain custom presets.
Users who already have the rest of LTX 2.3 can download only the new Foley and Licon LoRAs instead of downloading the full LTX bundle again.
Run Windows_Update_And_Open_Model_Report.bat afterward to compare the new preset requirements with the models currently on disk.

Fresh installation

Extract latest zip file into a clean folder.
Run Windows_Install_SwarmUI.bat.
Start the model downloader with Windows_Start_Download_Models_App.bat.
Download the bundle or individual models needed by your chosen presets.
Start SwarmUI and import latest Amazing_SwarmUI_Presets_.json
RunPod and Massed Compute users should follow the updated instruction files included in the package.

Massive ComfyUI Improvements

Furkan Gözükara — Thu, 27 Aug 2026 07:57:02 +0000

Zip file that includes everything with full changelogs shared here :

https://www.patreon.com/SECourses/posts/comfyui-auto-2-105023709
Video tutorial : https://youtu.be/1580ZDX-60Q

26 August 2026 Update V131

When you touch the prompt box, it was becoming shrinked and this issue resolved
MiniMax H3 Upgrades
- Continue videos using the final 1, 5, 22, 39, or 56 frames, instead of only the last frame.
- Multi-frame continuation uses MiniMax H3’s native clip guide for stronger motion, identity, and scene continuity.
- Replayed context frames and matching audio are automatically removed from the generated segment before merging.
- Continuation works with both FL2VA and Ref2VA without consuming an additional reference-image slot.
- The two Image To Video presets are now Image or Video To Video.
- So with these presets now you can do image to video or video to video

Start Frame / Init Video now accepts images or videos. A video can supply its final frame or selected multi-frame context.
Merge videos can append the new generation directly to the uploaded source video, preserving video order and audio.
Folder batches retain sequential generation, per-prompt saves, continuation, and final merged outputs.

Now we have contextual continue feature with Continue From Last Video Frames option we have

This feature implemented into SwarmUI as well so read it too
To update : get latest zip file, extract and overwrite all, then run Windows_Install_Or_Update_ComfyUI.bat and also Windows_Custom_Nodes_Bundles_Installer.bat (bundle 100) - all necessary custom nodes and extensions automatically installed and many of them managed by us for extra security

22 August 2026 Update V129

Python 3.10 not anymore supported because onnxruntime-gpu 1.27+ CUDA 13 build requires Python 3.11 or newer
- The installer will auto delete your existing Python 3.10 venv and auto generate new Python 3.12 venv if you have Python 3.12
- Please have Python 3.12 and use it
This is a very big update : 8 new official LTX 2.5 presets, MiniMax H3 Init Audio (make the video follow any soundtrack with lipsync), live token meter, Face Inpainting upgrades, automatic FL2VA / Ref2VA routing, a speed update and a big installer overhaul so please read all
8 new LTX 2.5 presets added

The 2-stage presets generate at 960x544 with 8 distilled steps, then apply the official LTX 2.5 latent spatial upscaler x2 and run 3 more refinement steps on a 1920x1088 canvas, and finally center crop to exact 1920x1080 - all with synchronized audio at 24 FPS and about 5 seconds by default
The Fast preset stays single stage at 960x544 for quick tests, and the First Last Frame preset is the official single-stage 1280x720 workflow : load a first and a last image and it generates the connected shots in between
The Expert preset exposes every stage on the canvas (Preprocess, Generate stage 1, Upscale and re-sampler stage 2, tiled Decode) with notes explaining each setting - the Text To Video and Image To Video presets are the compact subgraph versions, enter the subgraph when you want to fine tune

Audio To Video : load any audio, trim it with audio start and duration, and the video is generated to follow it (lipsync, action timing) - the encoded audio tokens stay frozen in both stages and your original trimmed audio is muxed into the output untouched, never decoded from latents - an optional start image is supported

Motion Track Control : drag points on the LTX Sparse Track Editor to sketch how things should move on your start image, and the motion track IC-LoRA animates exactly along those paths - a tracks preview video is saved next to the result so you can check the motion before committing to a long run

Text To Audio : audio only generation (sound effects, ambience, music cues) with the same LTX 2.5 model - no video VAE and no upscaler needed in this flow
Prompt enhancement (the official Gemma 4 E2B INT8 ConvRot enhancer) is off by default in every preset - enable it to expand short prompts into rich cinematic ones - and entering an LTX API key switches text encoding to the API to save VRAM
Models : one bundle for everything - open the latest model downloader and grab the LTX 2.5 Video Core Bundle : it now contains every model referenced by all ten of these LTX 2.5 presets : https://www.patreon.com/posts/114517862

LTX2.5 IC LoRA Pixel Spatial Upscaler presets fixed (FAST Distilled and Slower HQ) : the 4 helper nodes they need (LTX-2.5 Upscale Controls, Distilled Sigma Schedule, Prepare Video Chunks, Merge Video Chunks) are now part of our auto installed and managed by us extensions
Both upscaler presets also switched from the 24.46 GB BF16 text encoder to the 14.32 GB Gemma 4 12B INT8 ConvRot v2 one, so all ten LTX 2.5 presets now share the exact same text encoder and prompt enhancer - that is 10 GB less to download and one less model kept in memory when you switch between presets
New LTX2.5_Enchance_Prompt_Feed_For_LLMs.txt inside the Prompt_Generate_MiniMax_LTX_and_Demo_Materials folder : an official-aligned LTX 2.5 prompting guide for LLMs - give it to ChatGPT / Claude / Gemini, tell it what you want, and it writes properly formatted LTX 2.5 prompts for text / image / audio to video, multi-shot and IC-LoRA workflows
Optional Init Audio added to all six MiniMax H3 video presets (Text To Video, Image To Video, References To Video - both 8 Steps and 4x Speed) : select or upload any audio - or even a video file - in the new INIT AUDIO (OPTIONAL) node and the generated video follows that soundtrack exactly : lipsync, action timing, ambience - and keeps it as the exact final audio track

Leave it at (none - disabled) and the presets work exactly as before
Works with text only prompts, start frame and gallery references - by default the video duration automatically follows the audio length or set duration_mode to keep workflow duration
Init audio accepts video files with a visual trim panel - drag the handles to pick exactly which part of the soundtrack to use, nothing is re-encoded
Folder batches : put a same-name audio / video file next to each batch item and it is automatically used as that item's init audio (SAME-NAME INIT MEDIA)
Live MiniMax H3 token meter added to the reference gallery : prompt + references + resolution + duration are counted in real time against the 109k token budget so you see the usage before you queue - the meter also follows your init audio selection and trim window

Automatic FL2VA / Ref2VA routing : the MiniMax H3 video presets now pick the right checkpoint per prompt by themselves - single runs that have references and every folder batch item go through the Auto FL2VA / Ref2VA pipeline while plain text prompts keep the normal path - and the References To Video presets got an optional INIT IMAGE node : FL2VA uses it as the exact first frame, Ref2VA adds it as the starting frame picture reference (folder batch items ignore it and keep using their same-name init media and last frame continuation)

Face Inpainting upgrades : new faces to inpaint widget selects which faces get refined : 1 = biggest face (default), 2 = second biggest, 1,3 or all - faces are ranked biggest to smallest and each selected face is refined in its own pass and tagged in the tracking preview
New hallucination guard : every regenerated face crop is compared against its input and frames where H3 painted a neighbouring or different face over your subject are not pasted
Geometry lock is further improved and ON by default - removes the per-frame face shaking / tilting that the face pass could introduce
MiniMax H3 Text With References To Image preset also received the new face inpaint subgraph and a randomized seed

Speed : MiniMax H3 Sol-Attn sparse attention updated for SM89 (RTX 4000 series) and CUTLASS 4.7 in our maintained ComfyUI-TeaCache fork
Example on RTX 5090 : 1344x768 10 second video dropped from 450 seconds to 401 seconds with the updated speed optimizer
FLUX 2 Klein presets switched to my new INT8 ConvRot HQ models - almost BF16 quality but massively faster - auto download them with the latest model downloader : https://www.patreon.com/posts/114517862

Installer overhaul : custom nodes are now updated in-place instead of delete and re-clone so files downloaded inside node folders (for example ckpts) survive updates - a fresh clone replaces a node only when the in-place update fails
Leftovers of interrupted installer runs are cleaned automatically and staged copies use .disabled names so ComfyUI never imports duplicate nodes
Deleted venv is recreated in place without re-cloning ComfyUI - your models, outputs and user data are never touched - and the installer window now always stays open with a clear final status message
Python 3.10 support dropped because onnxruntime-gpu 1.27+ CUDA 13 build requires Python 3.11 or newer : menu now offers 3.11 / 3.12 (recommended) / 3.13
OpenCV conflicts between custom nodes fixed for good : all four OpenCV distributions are kept at one resolved version and opencv-contrib-python-headless is reinstalled last then verified - RES4LYF, controlnet_aux, Frame-Interpolation and rembg each pull different ones
onnxruntime-gpu upgraded to 1.27+ (CUDA runtime used by FantasyPortrait) and google-genai added to requirements
ComfyUI-Frame-Interpolation now installs without CuPy - it is not needed by the bundled RIFE workflows
RunPod and Massed Compute installers received the same in-place update and OpenCV normalization treatment
Windows_Run_GPU.bat renamed to Windows_Run_ComfyUI_Default.bat
To update : get latest zip file, extract and overwrite all, then run Windows_Install_Or_Update_ComfyUI.bat and also Windows_Custom_Nodes_Bundles_Installer.bat (bundle 100) - all necessary custom nodes and extensions automatically installed and many of them managed by us for extra security

16 August 2026 Update V122

Presets and custom nodes updated to improve face shaking effect that happens when automatic face inpainting activated
All presets now do have live preview (no increase in VRAM or generation duration) and fixed seed feature

16 August 2026 Update V120

MiniMax H3 References To Video 4x Speed - 24 FPS - 260808.json was broken with mistake and now fixed
Newest speed LoRA for MiniMax H3 is now in presets from LightX2V
Use latest Model downloader to get updated bundle : https://www.patreon.com/SECourses/posts/swarmui-auto-and-114517862

I have added automatic Face Inpainting feature to all MiniMax H3 presets
Get latest zip file, extract and overwrite all and run installer update bat file to update
Also use updated extra_model_paths.yaml if you are downloading models into SwarmUI

To understand how Face Inpainting works see this video clip
- https://www.reddit.com/r/comfyui/comments/1vpl5bb/secourses_comfyui_installer_and_ready_presets_now/
Live preview node added to all presets and does not cause any slowness or extra VRAM usage

14 August 2026 Update V117

Since you have requested, all MiniMax H3 presets modified and easy LoRA use added
- If you don't select any LoRA, workflow works just fine no issues so they don't break presets and how they work, they are optional

3 New presets added
- MiniMax Music 3 Text To Music - 260813.json - New model published today
- LTX2.5 IC LoRA Pixel Spatial Upscaler 2x - FAST Distilled 8 Steps - 260813.json
- LTX2.5 IC LoRA Pixel Spatial Upscaler 2x - Slower HQ - 260812.json

MiniMax Music 3 is reportedly lower quality than ACESTEP XL 1.5
- Our very advanced ACESTEP XL 1.5 Premium app below and it has full LoRA and DoRA training support
- App : https://www.patreon.com/SECourses/posts/ace-step-1-5-xl-157675060
- MiniMax Music 3 also missing necessary resources to be trainable
- MiniMax Music 3 prompt templates shared here : https://github.com/MiniMax-AI/MiniMax-Music3/tree/main/skills/music-caption-rewriter/templates
LTX 2.5 video upscaler preset is able to upscale any video, any resolution, any aspect ratio into 2x
- It is a very serious generative upscale so adds lots of details
- Therefore, try to upscale at 1 chunk to have better consistency
- Read preset information carefully after loading into ComfyUI
Our SECourses Premium Upscaler Pro app also fully supports LTX 2.5 now : https://www.patreon.com/SECourses/posts/secourses-pro-150202809
- It is standalone does not require ComfyUI and it auto downloads models when you first time use
To auto download necessary models of these presets, use our model downloader app and download from ComfyUI bundles : https://www.patreon.com/SECourses/posts/swarmui-auto-and-114517862

Moreover, I have converted the LTX 2.5 Int8 ConvRot myself higher quality and more accurate than officially published ones
- Someone detected one of the model is corrupt and broken - officially released ConvRot Int8 variant
- I converted them from BF16 versions after big research and experimentation
- So our model downloader app will download better quality Int8 ConvRot of LTX 2.5 models
Our Musubi Trainer convert to quant tab also now supporting LTX 2.5 Int8 ConvRot compile : https://www.patreon.com/SECourses/posts/secourses-musubi-137551634
LTX 2.5 video generation presets and bundles will be added soon hopefully but its quality lower than MiniMax H3
Get latest zip file, extract and overwrite all, then run Windows_Install_Or_Update_ComfyUI.bat and Windows_Custom_Nodes_Bundles_Installer.bat (select bundle 100) for updating all

12 August 2026 Update v116

To improve security, I have forked auto installed repos and bundle 100 repos and now they will be installed from secure fork
- Only few repos left as exception which are famous : mcmonkeyprojects , kijai , Lightricks
- I also did a massive deep security check of all auto installed and bundle 100 repos source codes and all are secure
MiniMax H3 presets got updated and now we have new features
Now when batch processing, you can dynamically set the duration of each prompt
- It will split the file name of the prompts with _ and if last split is a valid integer, it will be used as the duration of the prompt, otherwise whatever you set will be used
- Folder durations now use the final positive integer suffix only when the filename contains *: scene*7.txt → 7s; 200.txt, scene.txt, and scene_0.txt use the configured duration
- For example scene1.txt or 1.txt will use your set duration on the workflow meanwhile scene_1.txt will be set to 1 second duration dynamically
Added Continue from last frame beside Merge, disabled by default

Continuation automatically uses Ref2VA when references exist and FL2VA otherwise
Updated all MiniMax presets, extension UI, documentation, and the LLM prompt feed
- MiniMax_H3_Enchance_Prompt_Feed_For_LLMs.txt improved big time and includes latest changes and features as well
The most professional speed up LoRA maker LightX2V just published new speed LoRAs for the MiniMax H3
- https://github.com/ModelTC/Minimax-H3-Turbo

Our MiniMax H3 core bundles now auto downloads both FL2VA Turbo 8-step v1.0 and FL2VA Turbo 4-step v1.0 768p
- Use model downloader to auto download : https://www.patreon.com/SECourses/posts/swarmui-auto-and-114517862
All presets by default uses FL2VA Turbo 8-step v1.0 but you can switch it to FL2VA Turbo 4-step v1.0 768p
Older version 0.1 removed from downloads and presets so you can delete from LoRA folder if you downloaded
All presets exposes Shift value for Video and Audio seperately now
- So you can set to 6 / 3 if you switch to FL2VA Turbo 4-step v1.0 768p - default is 12 / 3
Sadly LoRAs only made for non-reference model atm but works on both models until they publish specific new LoRA for reference model
Some people asked me if video-latent continuation input is possible or not and sadly not possible so your best option is using continue from last frame
Get latest zip file, extract and overwrite all, then run Windows_Install_Or_Update_ComfyUI.bat and Windows_Custom_Nodes_Bundles_Installer.bat (select bundle 100) for updating all

11 August 2026 Update v113

RunPod and Massed Compute startup command changed so that the running ComfyUI instance won't be accessible by public IP anymore
- Updated files are : Massed_Compute_Instructions_READ.txt and RunPod_SimplePod_ComfyUI_Instructions_READ.txt
- Windows doesn't have such vulnerability it is fully locally running
Use Cloudflared secure HTTPs to access publicly when you are running on cloud services
- RunPod and Massed Compute installers auto installs Cloudflare
- Follow Massed_Compute_Instructions_READ.txt and RunPod_SimplePod_ComfyUI_Instructions_READ.txt
- Public access is optional, it will be default fully locally restricted running
When using unicode character containing references it was causing issues
- For example an image with Ü Ğ characters, this issue fixed and now works
Just run Windows_Install_Or_Update_ComfyUI.bat to update and always get latest zip file and overwrite older files

MiniMax H3 Infinite Video: Make 2-Hour Movies in ComfyUI (0-Shot)

Furkan Gözükara — Tue, 11 Aug 2026 02:30:55 +0000

Full tutorial link > https://www.youtube.com/watch?v=1580ZDX-60Q

Turn one folder of scene prompts into a long, coherent AI video with MiniMax H3 - locally, 0-shot and without babysitting every clip. This ComfyUI walkthrough shows how to match references, queue scenes, generate clips and automatically merge everything into one movie.
The opening is the raw workflow result. Then we rebuild it from installation to playback: models, presets, VRAM modes, prompt creation, folder batching, reference syntax, draft settings, troubleshooting, selective regeneration and consistency. It can scale to very long projects, including the 2-hour movie shown here.

COMFYUI INSTALLERS + PRESETS:

https://www.patreon.com/posts/comfyui-installers-presets-105023709

SWARMUI MODEL DOWNLOADER:

https://www.patreon.com/SECourses/posts/model-downloader-114517862

DISCORD HELP & SUPPORT:

https://discord.com/invite/software-engineering-courses-secourses-772774097734074388

Zip File Content Screenshots

Some Other Screenshots From Workflows

Ultra advanced multiple reference upload and mention system

Optional Low VRAM optimization with just a toggle enable disable

Trim and upload references feature

Automatically download necessary files ultra fast and SHA256 verified

Sana Labs published 4x speed up implementation

Video Chapters

[00:00:00] 0-shot generated movie showcase
[00:00:58] MiniMax H3 local workflow reveal
[00:01:10] Entire movies, prompts and scenes automated
[00:01:22] Audio references and folder-batch strategy
[00:01:40] Local desktop vs cloud deployment
[00:01:50] Download ComfyUI and the model downloader
[00:02:01] Fresh install and recommended Python 3.12
[00:02:17] Verify installation and install bundle 100
[00:02:37] Troubleshooting and support
[00:02:47] Final checks before launch
[00:02:57] Launch ComfyUI with run_gpu.bat
[00:03:11] 6 GB GPU support and speed tradeoffs
[00:03:22] Logs, errors and starting the preset
[00:03:33] Find the MiniMax H3 presets
[00:03:44] References-to-video 4x workflow
[00:03:54] Compilation, 20 steps and speed LoRA
[00:04:12] Change logs and Windows requirements
[00:04:24] Frequent workflow and custom-node updates
[00:04:36] Missing models and downloader setup
[00:04:48] Launch downloader and share model paths
[00:05:03] Core vs low-VRAM MiniMax H3 bundles
[00:05:16] INT4 vs recommended INT8 quality
[00:05:29] Options for 6-12 GB GPUs
[00:05:43] Select the correct models path
[00:05:55] Folder structure and model download
[00:06:07] Fix path warnings and select both VAEs
[00:06:23] Reference manager and default interface
[00:06:35] The enhanced prompt helper file
[00:06:47] Improve prompts with any major LLM
[00:06:59] ChatGPT setup with five voice attachments
[00:07:17] Example prompts and downloadable ZIP
[00:07:28] Prompt construction and easy referencing
[00:07:42] Upload the audio references
[00:07:52] Reference syntax and voice samples
[00:08:08] Listen to the supplied voice references
[00:08:18] Audio and image reference options
[00:08:28] Reorder image references by drag and drop
[00:08:38] Batch folder mode needs no manual prompt
[00:08:50] Set the folder path and draft resolution
[00:09:01] Aspect ratios and divisible dimensions
[00:09:15] Recommended 1344x768 for 16:9
[00:09:25] 30-second limit and 15-second sweet spot
[00:09:35] Best parameters: ready to run
[00:09:47] Run a quick 0.4 MP draft
[00:09:57] Automatic merging and queue monitoring
[00:10:08] Folder prompts and reference matching
[00:10:20] Load many references; use only the matches
[00:10:30] Per-generation reference limits
[00:10:40] Batch power for full videos and animations
[00:10:50] Iterate prompts with your favorite GPT
[00:11:00] Draft low resolution, review, then refine
[00:11:13] The key file for better prompts
[00:11:23] Single-clip mode and included presets
[00:11:33] Image, text and references-to-video presets
[00:11:46] Lightricks speed-up LoRA implementation
[00:11:59] LoRA version notes and future updates
[00:12:17] SwarmUI support and future advanced tutorial
[00:12:28] Read the docs and enable low-VRAM mode
[00:12:38] Save 15-40%+ VRAM
[00:12:54] Built-in node help and first output
[00:13:08] Regenerate only a weak section
[00:13:26] Play the generated result
[00:13:36] Current audio-reference limitation
[00:13:46] Improve voice and character consistency
[00:13:56] Match image IDs to characters
[00:14:06] Included reference and consistency guide
[00:14:16] 0-shot proof: ChatGPT authored all prompts, the movie script and rough draft
[00:14:35] Automated setup and broad GPU support
[00:14:45] Final requirements reminder and outro
MiniMax H3 supports 30-second segments; 15 seconds is the sweet spot. This demo uses 10-second scenes for easy review and regeneration.
For 16:9, use 1344x768. Draft around 0.4 MP, review the sequence, improve prompts, then render final quality.
Load up to 99 references; each prompt uses only matched IDs. Per generation: up to 3 audio, 3 video and 9 image references. Extra audio/video refs are discarded.
Use INT8 for speed and near-BF16 quality; INT4 is for lower VRAM. It runs on 6 GB GPUs, but more slowly. Low-VRAM mode saves about 15%, reaching 40%+ in some cases.
Audio references are less reliable than image references. If voice quality drops, try no audio reference. Map image IDs clearly for consistency.
Covers local Windows, Massed Compute, RunPod and SimplePod. Use a fresh install with Python 3.12 where recommended. Read requirements and changelogs; nodes and workflows update often.
If one scene fails, regenerate only that section. Join the Discord above for setup help.

MiniMaxH3 #ComfyUI #AIVideo #LocalAI #GenerativeAI

Ultimate Image Processing Toolkit by SECourses

Furkan Gözükara — Tue, 28 Jul 2026 17:02:10 +0000

Ultimate Image Processing APP Premium by SECourses : Batch Cropping, Zooming In, Resizing, Duplicate Image Removing, Face Extraction, SAM 2, SAM 3.1 and Yolo Segmentation, Masking for Windows, RunPod, Massed Compute

Installers and APP is shared here with instructions : https://www.patreon.com/posts/120352012

APP : https://www.patreon.com/posts/120352012

Video Tutorial : Hopefully coming soon

Batch Image Cropping, Zooming Subject, Resizing, Segmenting, Masking, Duplicate Removing APP that utilizes YOLO V26, YOLO Face V12, SAM 2, SAM 3.1 with 1-click installers for Windows, RunPod, SimplePod and Massed Compute (Linux users use this)

28 July 2026 V1.0

We have completey remade the app and read below to understand how it works and full features
Make a fresh install and read the below carefully please
We use latest pre-compiled wheels and Torch 2.13 and CUDA 13

Windows Requirements

Python 3.12.10, FFmpeg, CUDA 13, cuDNN 9.17 or above, Visual Studio Community Edition with all C++ options selected
- Don't worry CUDA 13 works with all GPUs - make sure you have updated NVIDIA driver
- Follow this requirements tutorial video exactly : https://youtu.be/DrhUHnYfwC0
- Follow its updated post with links and screenshots exactly : https://www.patreon.com/SECourses/posts/requirements-written-tutorial-111553210

Turn a folder of raw photos into a clean, correctly framed, ready-to-train dataset. The nine tabs are designed to feed each other: the output of one is exactly what the next expects to read.

One architectural detail matters more than any single feature: the Gradio process never has to import torch. Tick Run every task in a subprocess on the Performance tab and every job runs in a throw-away process that loads the model, does the work and exits, returning 100% of its VRAM, RAM and CUDA context the moment it finishes. A long session never creeps upwards, and the app sits at zero idle VRAM between runs. It ships off, so jobs run in-process by default and keep their models resident for faster repeated runs — turn it on when you would rather have the memory back.

Requirements
Installation
Launching the app
Presets
Tab 1 — Image Cropper
Tab 2 — Segmentation
Tab 3 — Extract Faces & Objects
Tab 4 — Image Resizer
Tab 5 — Generate Tiled Images
Tab 6 — Find Duplicates
Tab 7 — Move Low-Res Files
Tab 8 — Rename Files
Tab 9 — Performance & VRAM
Recommended workflow
Output folder layout
Measured benchmarks

Requirements

Python 3.12.10 (3.10.x / 3.11.x / 3.13.x may also work)
CUDA 13.0 and cuDNN 9.17+
Git, FFmpeg
Visual Studio Community Edition with all C++ options (needed for torch.compile)
An NVIDIA GPU — any card works, more VRAM is better. CPU-only is supported but slow.

Setup walkthrough for the prerequisites: https://youtu.be/DrhUHnYfwC0

Installation

Windows — automatic

Run Windows_Install_or_Update.bat. It clones this repository, creates a Python 3.12 venv, installs the dependency set with uv, and then downloads every model via HF_model_downloader_img_process.py.

Re-run the same script later to update — it does a git reset --hard + git pull before reinstalling.

Manual

git clone --depth 1 https://github.com/FurkanGozukara/Ultimate_Image_Processor_Pro
cd Ultimate_Image_Processor_Pro
py -3.12 -m venv venv
call .\venv\Scripts\activate.bat
python -m pip install --upgrade pip
pip install uv
cd ..
uv pip install -r requirements_image_process.txt --index-strategy unsafe-best-match
python HF_model_downloader_img_process.py

The requirements file pins torch==2.13.0+cu130 plus prebuilt wheels for xformers, FlashAttention, SageAttention, torchao, InsightFace and SAM 2 — so no compilation is needed on a supported platform.

Cloud (RunPod / SimplePod / Massed Compute)

See the instructions bundled with the Patreon release.

Launching the app

Use Windows_Start_App.bat, or activate the venv and run:

python app.py

Flag	Effect
`--share`	create a public Gradio link
`--server-port PORT`	serve on a specific port (default: auto)
`--server-name HOST`	host to bind to (default: `127.0.0.1`)
`--no-browser`	do not open a browser window on launch
`--quiet`	reduce Gradio's start-up output

Presets

A preset captures every control in the entire app, across all nine tabs — around 150 of them. The preset you last saved or loaded is restored automatically the next time the app starts, so your working configuration survives a restart.

Selecting a preset in the dropdown changes nothing until you press Load, so a mis-click can never wipe the settings you are in the middle of editing. The built-in Default set lives in code rather than on disk, so it can never be overwritten or deleted.

Tab 1 — Image Cropper

Finds the subject in every photo of a folder and crops the tightest box that matches each aspect ratio you ask for. It detects the subject, adds your padding, then grows the box outwards until the ratio is exact — so the subject is never squashed and no part of it is ever cut off.

Output goes to one sub-folder per aspect ratio (1024x1024/, 1280x720/, 1080x1350/ …), which is exactly the layout the Image Resizer tab reads back.

Live folder scan — image count, extensions and total size shown before you press start
Three detection engines, each showing only its own thresholds
All instances (union) — spans every match so nobody is cropped out of a group photo
Padding in percent or pixels, applied before the ratio so the ratio stays exact; negative values crop in tighter
Annotated previews — optionally writes a copy with detections drawn on it, to check a prompt before a big run
Safe re-runs — with overwrite off, images whose crops already exist are skipped

The three detection engines

Engine	What it is	Its knobs	Use when
SAM 3.1 (default)	Detects and segments in one model from open-vocabulary text. Pixel-accurate masks, no fixed class list.	confidence, max instances	your subject is not a COCO class, or you need a real mask
YOLO	Fastest, no prompt wording to get wrong. 80 COCO classes plus a dedicated face model (`yolov12l-face.pt` swaps it in automatically).	confidence, NMS IoU, inference size, max detections	your subject is a COCO class and you want throughput
SAM 2.1 + GroundingDINO	The older two-stage pipeline: GroundingDINO proposes boxes, SAM 2.1 turns them into masks.	box threshold, text threshold	reproducing results from before SAM 3.1

How the cropping logic works

Detect → pad → grow outwards to the ratio. And the detail that surprises most people: nothing is ever upscaled. The aspect ratio you type is used as a ratio, not as a target size, so a 6000 px wide photo yields a 6000 px wide 16:9 crop, not a 1280 px one. Tick Resize crops to the exact target size if you want the Cropper to scale too — otherwise crop here and resize in the Image Resizer tab, where you choose the filter. That order throws away the least detail.

The run panel

Every batch tab shares the same panel: a live status card with progress, images/sec and ETA, a copyable log naming each file and how many outputs it produced, and a Stop button that really stops the worker and keeps whatever was already written.

Real results

The annotated detection preview alongside the 1:1, 16:9 and 4:5 crops of the same photos. Every crop is a plain rectangle of original pixels — the app never resamples during cropping.

Tab 2 — Segmentation

SAM 3.1 masks from plain English, with an interactive single-image tester and a folder batch that share one set of settings.

Dial your prompt in on one image first — it takes a few seconds and shows the exact mask, the confidence score and the mask area in pixels. The batch section further down the page then uses those identical settings; there is no second set of controls to keep in sync.

You get the segmented overlay, the mask itself (white inside, black outside — ready for inpainting, ControlNet or masked training), and a copyable detections table with label, score, bounding box and mask pixel count per instance.

Negative prompts

Subtract one concept from another in a single pass. Below, the prompt is face. and the negative prompt is eye. lips. — the eyes and mouth are carved straight out of the face mask. The negative mask is saved separately so you can reuse or invert it.

This is how you build skin-only masks, or segment a car while cutting out the person standing in front of it.

Settings and the YOLO backend

Concepts are separated by periods and the vocabulary is open. Confidence drops weak instances; max instances caps both the result count and how much VRAM the mask upsample needs on a busy scene; overlay opacity affects only the preview tint (the saved mask is always pure black and white).

On the YOLO backend you multi-select any number of the 80 COCO classes instead of writing a prompt, with a separate exclude list. YOLO mask source decides where masks come from: SAM 2.1 refined turns each box into a pixel-accurate mask (loads a second model, still fast), while filled boxes is instant and plenty for rectangular crops or inpainting blocks.

Batch processing

One mask per image, plus optional annotated overlays and negative masks. Leaving the overlay folder empty makes a run noticeably faster, because the masks then never need upsampling to full resolution for display. Skip images that already have a mask makes an interrupted run cheap to resume.

The prompt below was a single word: face. The bottom row is a red-carpet shot with a crowd in the background — SAM 3.1 found 16 separate faces in it and every one is in the mask.

Tab 3 — Extract Faces & Objects

Cuts the detected region out of each photo and saves it as its own image — the fastest way to build a face dataset. Unlike the Image Cropper this does not force an aspect ratio: you get the detection box plus whatever padding you ask for.

Save every detected instance as its own file — a group photo of five people yields five crops (photo_1.png … photo_5.png). Off gives exactly one crop per photo, chosen by confidence or largest area.
Padding — faces almost always look better with 15–30% of breathing room. Percent scales with the face, so close-ups and wide shots come out consistent.
Minimum crop size — one slider that drops background faces too small to be useful.
Not limited to faces: hand. licence plate. dog. all work.

Tab 4 — Image Resizer

Writes every input image at one or more exact pixel sizes. Each input is written once per resolution into <output>/<WxH>/, so several dataset buckets come out of a single run.

Run it after the Image Cropper: crop first at native resolution, then resize once with the filter you want.

It works out what you pointed it at

Point at an Image Cropper output folder and it reports Image Cropper layout detected, listing how many files each WxH bucket contributes and naming any resolution with no matching sub-folder. Point at a plain folder of pictures and it says Plain folder detected and does the multiplication for you: 12 images × 3 resolutions = 36 output files.

Framing choices

Crop to fill keeps the aspect ratio and trims the overflow — no bars, some content lost, and the default. Fit inside keeps every pixel and pads with your chosen background colour. Stretch distorts the image and is almost never what you want.

Separately, Centre the crop on decides where the crop is taken. Image centre is instant and loads no model, but a geometric centre crop can push the subject out of frame. Detected face / subject runs a detector first so the subject stays in frame — slower, since each worker loads its own model, so keep the worker count low in that mode.

Tab 5 — Generate Tiled Images

Slices each photo into an overlapping N×N grid of fixed-size tiles. Because the tiles are cut at native resolution and never upscaled, this is a very good way to turn a handful of very large photos into a high-detail training set.

Tiles per image must be a perfect square — 9 means a 3×3 grid — and tiles overlap automatically so the whole frame is covered. If an image is smaller than one tile you choose the policy: Skip protects quality, Upscale guarantees output but invents pixels, Shrink the tile keeps the grid but writes smaller tiles for those images only.

Tab 6 — Find Duplicates

Groups visually identical or near-identical images and moves the extras out of the way, keeping one copy per group. Every comparison uses a true perceptual distance — earlier versions compared hashes as big integers, which is not a similarity measure at all and both missed real duplicates and invented fake ones.

Nothing is deleted. Duplicates are moved to a destination folder keeping their relative structure, and a full report is written next to them, so the step is reversible. You choose which copy survives: highest resolution, largest file, oldest, or first alphabetically.

Choosing an algorithm

The app ships a reference table for aHash, dHash, pHash, wHash, ColorHash, KAZE local features and a combined multi-hash mode — what each measures, what it is good at, where it breaks down, and the cutoff to start from.

Cutoffs are Hamming distances: how many bits differ out of hash size², so smaller is stricter. Each algorithm gets its own threshold, because a distance of 5 means something completely different for aHash than for pHash. KAZE is the odd one out — it survives rotation, perspective and heavy cropping, but it is quadratic in image count, so keep it for small folders.

There is also a safety cap on image count, because comparison is quadratic and a 100,000-image folder would otherwise run for hours. Anything dropped by the cap is reported explicitly rather than silently ignored.

Tab 7 — Move Low-Res Files

Quality control for a crop run. Any processed image smaller than your threshold is moved out of the way together with the original it came from, so your two folders stay in sync — you never end up with a source photo whose result has been thrown away.

Set the minimum width and height, then choose whether to reject when either dimension is under the limit (strict) or only when both are. Files are paired by name, and Same name, any extension handles the everyday case where the crop is a .png and the original a .jpg. Preview mode is on by default and lists exactly what would move without touching disk.

Tab 8 — Rename Files

Renames a folder to a clean prefix + number sequence. The rename happens in two passes through temporary names, so a target name already taken by another file in the same folder can never clobber it. Extensions are always preserved.

Set the prefix, the starting number, the zero padding (4 or 5 so files sort correctly everywhere) and the numbering order — by name, by date, or by size. Only rename image files leaves captions and JSON sidecars alone. Preview mode is on by default and shows the first 50 planned renames without touching disk.

Tab 9 — Performance & VRAM

Everything on this tab applies to every other tab and is saved with your presets. The shipped defaults are already the fastest configuration that still matches the reference model output — you only need this tab to trade quality for memory, or to try torch.compile.

Run every task in a subprocess (off by default, tick it if you want it) — each run starts a fresh Python process, works, then exits, returning 100% of its VRAM and RAM, CUDA context and all. Left off, models stay resident between runs: faster repeated runs, but the VRAM stays held.
Switching model frees the old one first. Only one model set is ever resident. Change the backend, the precision, the resolution, the GPU or the YOLO class that swaps in the face weights, and the previous model is unloaded — collected out of RAM and released from VRAM — before the new one is read from disk, so a switch never has to hold both. Measured on an RTX 5090: SAM 3.1 (1861 MB) → YOLO leaves 262 MB held, and eight switches back and forth drift by 0 MB. The console names what it dropped: [models] released SAM 3.1 (concept prompts) [bfloat16 @ cuda:0, res 1008]. If a second task happens to be running when you switch, its model is left alone until the process is quiet again.
Unload models and release VRAM — the diagnostics button below drops every loaded model on demand and reports the before/after figures, e.g. model weights 1893 MB → 0 MB, allocator reserve 2670 MB → 0 MB. What is left after that is the CUDA context itself, which only a process exit returns.
GPU IDs — comma-separated CUDA indices. With several GPUs the image list is split across them, but each device loads its own copy of the model, so a second GPU only pays for itself on folders of a few hundred images or more. Enter cpu to force CPU.
Model precision — the shipped SAM 3.1 checkpoint is stored in bfloat16, so keeping the weights in bfloat16 is bit-identical to the published values. Measured against the float32 path: 0.9992 mask IoU, 46% less VRAM (1.9 GB vs 3.5 GB), ~7× faster model loading and ~5% faster per image.
Load weights straight into VRAM — builds the model on the GPU and reads the safetensors file directly into VRAM, skipping the ~3.4 GB float32 copy in system RAM. Roughly 2.5× faster model loading.
Chunked mask upsampling — upsamples instance masks a few at a time. On a 24 MP photo with 20 instances this is the difference between a ~1 GB spike and a few tens of MB, with pixel-identical results.
Attention kernel, TF32, cuDNN autotuner, channels-last, matmul precision — all exposed and documented inline.
torch.compile the vision backbone — a one-off compile cost of roughly one to three minutes (cached under .cache/uipp) for faster steady-state inference. Worth it for large batches, not for a handful of images; falls back to eager mode if compilation fails.
SAM 3.1 inference resolution — 1008 is what the model was trained at. Lower is faster but loses small objects; higher costs quadratically more and adds no detail the model was trained to use.

Runtime probe

One click imports torch — in a throw-away subprocess if Run every task in a subprocess is ticked, otherwise in this process — and reports what your machine can actually do: Python and platform, torch and CUDA versions, every GPU with its compute capability and VRAM, the status of Triton / FlashAttention / xformers / SageAttention / torchao, whether torch.compile is genuinely usable on this box (CUDA root, host compiler, ninja, cache path), and whether each model backend is ready. In subprocess mode the process then exits and the memory is gone.

Recommended workflow

Find Duplicates — clean first, so you never spend GPU time on near-identical shots
Image Cropper — frame the subject, one sub-folder per ratio, at native resolution
Move Low-Res Files — quality control, results and originals moved together
Image Resizer — point it at the Cropper's output folder for exact pixel sizes
Rename Files — a clean, correctly sorting sequence

Side branches whenever you need them: Segmentation for masks, Extract Faces for a face dataset, Generate Tiled Images for more data out of very large photos, and Performance & VRAM set once for everything.

The one rule that matters most: crop before you resize. Cropping keeps the original pixels, so resizing afterwards starts from the largest possible image and only has to downscale once. The other way round throws away detail you can never get back, and can force an upscale.

Output folder layout

The Image Cropper's <width>x<height>/ sub-folder layout is not cosmetic — the Image Resizer reads it directly, matching each bucket to the resolution of the same name. That is exactly why the recommended order is crop → resize.

Quick reference

Every tab, the one setting that matters most, and the mistake to avoid.

Measured benchmarks

Single RTX 5090, 12 source photos at up to 6714 × 5274 px, SAM 3.1 at its default settings.

Job	Work done	Time
Image Cropper	12 photos → 36 crops across 3 aspect ratios	47.7 s (0.25 img/s)
Segmentation (batch)	12 photos, 33 instances masked	28.8 s (0.42 img/s)
Extract Faces	33 face crops from 12 photos	9.6 s (1.25 img/s)
Image Resizer	36 files at exact pixel sizes	2.0 s (17.7 img/s, CPU)
Generate Tiled Images	108 tiles from 12 photos	1.2 s
Find Duplicates	4 groups across 17 images	0.4 s

Built by SECourses · Patreon · YouTube

Ideogram 4 Is Here: The Ultimate JSON Prompting Masterclass

Furkan Gözükara — Sun, 19 Jul 2026 13:30:16 +0000

Learn how to run Ideogram 4 locally with SwarmUI and ComfyUI, download the required model bundle, use the ready Turbo, Balanced, and Highest Quality presets, and create accurate structured JSON prompts with Ultimate Image Captioner Pro.

The workflow covers image recreation, reliable text rendering, bounding-box editing, batch captioning, and training-dataset preparation.

Full Tutorial

Watch the complete step-by-step tutorial on YouTube:

Ideogram 4: The Ultimate JSON Prompting Masterclass

Tutorial Resources

Resource	Link
Ultimate Image Captioner Pro download and installer	Patreon post
SwarmUI installer, model downloader, and presets	Patreon post
ComfyUI installer	Patreon post
Windows requirements tutorial	YouTube
Requirements guide with links and screenshots	Patreon post
Community support	SECourses Discord
More tutorials, installers, and resources	SECourses GitHub

Supported Models

Ultimate Image Captioner Pro supports the following models with robust torch.compile integration.

Qwen Vision Models

Qwen3-VL 8B Instruct (default)
Huihui Qwen3-VL 8B Instruct Abliterated
Qwen3-VL 4B Instruct
Qwen3-VL 2B Instruct
Qwen3-VL 30B-A3B Instruct
Qwen3.6 27B
Huihui Qwen3.6 27B Abliterated

Joy Caption Models

Joy Caption Beta 1
Joy Caption Alpha 2
Joy Caption Alpha 1
Joy Caption Pre Alpha

Torch 2.13 Runtime

The application uses Torch 2.13 with the latest project-tested, precompiled supporting libraries.

Application Preview

Click the image to open the full-size screenshot.

Torch Compile Performance

The fully compiled captioning path delivers an 84% speed improvement in the demonstrated benchmark.

Installers

Installer workflows are available for Windows, RunPod, SimplePod, Massed Compute, and local Linux systems.

Video Chapters

Show all tutorial chapters

00:00:00 - Ideogram 4 overview: JSON prompting, SwarmUI presets, ComfyUI workflows, and model bundle
00:00:53 - Ultimate Image Captioner Pro for turning reference images into Ideogram JSON prompts
00:01:10 - Editing JSON elements, bounding boxes, wanted text fields, captions, and prompt layout
00:02:02 - Regeneration examples showing structure, objects, scene layout, and image text matching
00:03:18 - Captioner Pro feature tour: Qwen, JoyCaption, saved outputs, and JSON builder
00:04:30 - Dataset workflow: prompt presets, batch folder captioning, and automatic VRAM presets
00:05:13 - Tutorial roadmap: ComfyUI update, SwarmUI update, model download, app installation, and usage
00:05:40 - Updating ComfyUI by extracting the latest installer ZIP and overwriting old files
00:05:56 - Optional fresh ComfyUI virtual environment rebuild for outdated or broken installations
00:06:15 - Running the ComfyUI update script, Python choice, UV speed, and quantization support
00:07:05 - Installing recommended custom nodes bundle 100 for ComfyUI and SwarmUI compatibility
00:07:50 - Launching fresh ComfyUI and testing the Ideogram Turbo preset workflow
00:08:47 - Setting width, height, resolution, and matching the prompt aspect ratio
00:09:07 - Updating SwarmUI with the latest ZIP, overwrite method, and safe folder paths
00:09:48 - Automatic .NET SDK 10 installation and why SwarmUI needs the correct SDK version
00:10:51 - SwarmUI backend setup: ComfyUI backend, Triton, Sage Attention cautions, and extra arguments
00:11:44 - Downloading the Ideogram 4 core bundle with hash verification
00:12:28 - 16-connection parallel downloads, target folders, ComfyUI mode, and URL downloader
00:13:20 - Merging model parts and sharing SwarmUI models through extra_model_paths.yaml
00:13:51 - Setting the SwarmUI model root to reuse another model folder and avoid duplicates
00:14:12 - Updating SwarmUI presets with delete import, normal import, overwrite, and backup
00:14:58 - Refreshing presets and confirming Ideogram Turbo, Balanced, and Highest Quality
00:15:14 - First simple Ideogram prompt, false safety-filter block, and weak plain prompting
00:15:34 - Using Realism Engine Ideogram 5 LoRA to fix the blocked car prompt
00:15:57 - Why detailed JSON prompts are needed and downloading Captioner Pro
00:16:23 - Installing Captioner Pro with the Windows install/update app, virtual environment, and model downloads
00:16:34 - Windows requirements: Python, CUDA, cuDNN, C++ tools, FFmpeg, Git, and setup guide
00:17:03 - Cloud and Linux notes plus the Massed Compute interface, creator image, GPU, and coupon
00:17:34 - Captioner installer downloader: 16 connections, hash checks, and accurate setup
00:17:57 - Starting Ultimate Image Captioner Pro and saving custom user presets
00:18:14 - Loading the Bugatti reference image and generating official Ideogram JSON
00:18:39 - Prompt generation speed, copying the prompt, and understanding VRAM usage
00:19:09 - Subprocess mode to release all VRAM and RAM after each captioning run
00:19:54 - Reviewing generated JSON: high-level description, visible text, boxes, and details
00:20:21 - Pasting JSON into SwarmUI and matching the custom 5:3 aspect ratio
00:20:43 - Aspect-ratio calculator, side-length control, and high-resolution generation
00:21:36 - Comparing results with and without aspect-ratio metadata and avoiding false safety blocks
00:21:58 - Realism Engine LoRA strength, when to use it, and output comparison
00:22:34 - Choosing Turbo, Balanced, or Highest Quality and testing Turbo speed
00:22:54 - Ideogram 4 image-to-image, inpainting, image creativity, and image prompts
00:23:19 - Captioner Pro batch-folder processing: subfolders, overwrite, and append modes
00:23:35 - Post-processing captions with prefixes, suffixes, replacements, and sensitivity
00:24:07 - Final options, automatic quantization by GPU VRAM, support channels, and closing

Covered in the Tutorial

Local Ideogram 4 installation
SwarmUI and ComfyUI preset usage
Automatic model downloads and hash verification
Structured JSON prompt creation
Bounding-box and visible-text editing
Reference-image recreation
Safety-filter troubleshooting and LoRA realism settings
Folder-based batch captioning
VRAM-friendly caption generation

ACE-Step XL 1.5 Premium + Facebook / META SAM Audio + Auto-Editor Trim + Audio Tools Enhancement Full Tutorial

Furkan Gözükara — Sun, 28 Jun 2026 00:59:23 +0000

ACE-Step XL 1.5 Premium + Facebook / META SAM Audio + Auto-Editor Trim + Audio Tools Enhancement Full Tutorial

Video Tutorial

https://youtu.be/9C_6qNKjgpA

Source Video, Links, And Chapters

The public video description presents this as a full ACE-Step XL 1.5 Premium guide for local AI music generation, remix, repaint, stem extraction, wildcard prompt variation, audio processing, SAM Audio segmentation, Windows installation, RunPod, Massed Compute, SimplePod, and Linux/cloud workflows.

Tutorial video: Forget Suno: Run the Ultimate AI Music Studio LOCALLY (100% Free)
Channel: SECourses
App/latest zip and Patreon post: Patreon ACESTEP XL Premium SAM Audio post
Windows requirements guide: Windows requirements guide
Discord/help/community: SECourses Discord server
Duration: 1:15:09
Published: June 17, 2026

Video Chapters

0:00 — Intro: ACESTEP XL 1.5 Premium local music, segmentation and processing tutorial
0:52 — Fast song generation examples across styles in under one minute
1:55 — Output manifest proof, 40-second generation time and supported models
2:29 — Turbo/SFT/Base models, LoRA support, GPU presets and Torch Compile boost
3:10 — Remix feature preview, same-lyrics requirement and responsible usage note
4:16 — Repaint mode: regenerate and merge only a selected song section
5:38 — Extract mode: stems, silence trimming, all-stems and batch folders
6:30 — Lego mode: add an instrument stem such as guitar into existing audio
7:25 — Audio Processing presets and manual enhancement controls for AI songs
8:35 — Auto-Editor silent trim for tutorials, videos, audio and workflow export
9:48 — DaVinci/Premiere/Final Cut/ShotCut/Kdenlive timeline export demo
11:01 — SAM Audio Segment: BF16 models, VRAM presets and advanced segmentation
11:47 — SAM outputs demo: vocals, drums, bass, remaining audio and saved files
12:47 — Custom SAM prompts, semicolon batch segmenting and speech cleanup example
14:19 — Batch processing, load metadata, manifests, saved settings and presets
15:09 — Why local open-source models matter and where to run ACESTEP
15:55 — Windows install begins: Patreon zip, changelog, attachments and download
16:53 — Windows requirements tutorial before Python/CUDA/C++/FFmpeg setup
17:29 — Extract zip safely, avoid bad paths and run Windows_Install_or_Update.bat
18:24 — Automatic VENV, FFmpeg, UV install, model downloads and hash verification
19:24 — Turbo default vs all-model download for SFT/Base and BF16 safetensors
20:32 — First Windows launch, default Generate Song test and CMD progress
21:44 — Model recommendations, VRAM tiers, languages, vocals and MP4 image output
23:29 — Torch Compile setup for faster repeated generations
24:05 — Outputs folder, model switching and full remix setup workflow
25:24 — Practical remix loop: adapted lyrics, strength, reference audio and seed lock
28:03 — Repaint workflow with source range preview, generated result and comparison
29:13 — Recap: extraction, Lego, audio processing and SAM text-prompt usage
30:20 — Windows wrap-up, LoRA training teaser and move to cloud installs
31:16 — RunPod setup: credits, template, CUDA filters, GPU choice and storage
34:53 — Upload zip in Jupyter Lab, extract, run instructions and handle installs
35:43 — RunPod errors, resume behavior, model downloads and hash verification
38:04 — Start ACESTEP on RunPod with Gradio Live, proxy ports and persistence
40:18 — Add 7860/7861 ports, verify storage reuse and rerun installer after resume
42:10 — RunPod connection troubleshooting and Gradio Live recommendation
44:12 — Fix corrupted VENV/stale handle errors, reinstall safely and retest
47:24 — Successful RunPod relaunch, default generation, nvitop and loading tips
49:26 — RunPod first load vs fast inference, 15-second second generation example
51:02 — Download outputs and delete RunPod pods/storage to stop spending
53:30 — Massed Compute setup: coupon, Creator image, GPU prices and ThinLinc
57:13 — Massed install from extracted folder, Linux notes and ultra-fast downloads
59:18 — Start app on Massed Compute via localhost or Gradio Live
1:00:23 — Default Massed generation, nvitop, faster loading and speed test
1:02:03 — Sync/download outputs and delete Massed Compute instance safely
1:03:25 — SimplePod setup: template, persistent volume, pricing and GPU choice
1:06:39 — Jupyter upload, direct file browser, install command and model downloads
1:08:21 — Start SimplePod, Gradio Live, default generation and one-time load errors
1:09:31 — nvitop monitoring, newer driver/CUDA details and generation completion
1:10:42 — Direct output/model downloads through SimplePod file browser
1:11:42 — Delete instance, keep storage, relaunch GPU and verify install
1:13:15 — Discord, subreddit, changelog, update guidance and support links
1:14:30 — Final cleanup: terminate servers, delete storage and LoRA training outro

Few Updates Made Since Video Tutorial So They Are Listed Here

18 June 2026 V5.5 Update

Full tutorial video published finally for inference : https://youtu.be/9C_6qNKjgpA
I started working on LoRA training tutorial as well hopefully soon
With 5.5 optimizer specific parameters are now shown that you can set, I am also working on to make them auto default best hopefully
There was a visual bug that hidden Remix Melody Retention and Direct Source Latents (no_fsq) on Remix songs page and this bug fixed and app scanned entirely and all visuals verified
- Default value set to 0.97 one of our expert remixer recommended that
Just run Windows_Install_or_Update.bat to update, the zip file not changed

18 June 2026 V5.4 Update

Now batch folder processing for ACESTEP XL 1.5 and SAM Audio has this extra option Save only output
- This is useful to get only processed files and no other stuff like remaining part of the songs or metadata files, etc.

18 June 2026 V5.3 Update

Wildcard feature implemented
It works both for Style / Captions and Lyrics with syntax verification as well
It will work in batch folder processing as well so you can write that way in txt files
- If you enable Auto improve lyrics or Auto improve style they may break your syntax so don't enable when using wildcards
Also full inference tutorial published that covers every topic in details including how to install on Windows, RunPod, Massed Compute and SimplePod : https://youtu.be/9C_6qNKjgpA

1. What ACE-Step XL 1.5 Premium Is

ACE-Step XL 1.5 Premium is a local-first music generation and audio utility suite. The video presents it as more than a song generator: it also includes Wildcards for prompt variation, advanced generation modes, remix/repaint workflows, stem extraction, LEGO-style stem addition, SAM Audio segmentation, Auto-Editor trimming, mastering-style audio processing, dataset tools, and LoRA/LoKr training pages.

Responsible-use note: the source tutorial says to use the application respectfully and for research/education. For remix, repaint, extraction, and pitch work, use material you own, have permission to process, or are otherwise allowed to use.

Video introduction

Feature overview

Core jobs covered in the tutorial:

Generate complete songs from a style prompt, structured lyrics, model choice, duration, language, and seed settings.
Use Wildcards in style, Music Caption, and Lyrics fields to randomly pick bracketed options at generation time.
Remix, repaint, extract, LEGO-add, complete, retake, edit, and reuse LM code hints when the selected model supports the workflow.
Trim silence, export audio/video, enhance or pre-master generated songs, and optionally run pitch correction with DiffPitcher.
Use SAM Audio with quick prompts, custom prompts, explicit spans, or batch prompt lists to extract target audio and save the remaining audio.
Use Library, Load Metadata, Results, and the preset system to restore, inspect, score, save, and reuse generation runs.

2. Install And Start On Windows

The Windows workflow uses the included batch files. Extract the ZIP, keep the folder structure intact, run the installer/update script, optionally download all models, then start the app with the Windows launcher.

Windows installer

Extract the ACE-Step Premium ZIP to a path with enough free disk space for the virtual environment, model files, outputs, and FFmpeg runtime.
Run Windows_Install_or_Update.bat. The installer creates the Python virtual environment, downloads or uses shared FFmpeg, installs packages with UV, and prepares the app.
Run Windows_Download_All_Models.bat if you want SFT and Base models in addition to the automatically available Turbo path.
Run Windows_Start_App.bat. In this workspace the launcher started ACE-Step at http://127.0.0.1:7862 because other Gradio apps were already using 7860 and 7861.
Watch the command window for model download, model load, generation, and error details. The video recommends trusting the terminal status more than only the browser UI.

Model availability: Turbo is the quick default. SFT and Base require additional model files. Remix is recommended with SFT in the video; some modes are marked Base-only or unavailable until the matching model is selected.

Windows first generation

3. Quick Song Generation

The Generate Song tab is the fast path. It exposes the controls most users need: style, lyrics, Wildcards, model, LoRA, GPU preset, quantization, language, vocal type, instrumental toggle, duration, count, seed, optional MP4 image, and video resolution.

Generate Song overview

Generate Song filled

Wildcards For Prompt Variation

ACE-Step XL 1.5 Premium v5.3 adds Wildcards. Write bracketed choices separated by pipes, such as [option A|option B|option C], and one option is picked when you generate. Wildcards can be used in the quick Generate Song Style field, the Advanced Music Caption field, and Lyrics.

Wildcards in Generate Song

Wildcards in Advanced caption and lyrics

Basic syntax: [piano|guitar|synth] picks one option when the job starts.
Nested syntax is supported. Example: cinematic [piano|guitar [clean|crunchy]] hook.
Lyrics can use the same pattern, for example: I feel [alive|ready|free] tonight.
Lyric tags without a pipe, such as [Verse], [Chorus], and [Instrumental], stay unchanged.
Do not enable Auto/Enhance Style or Auto/Enhance Lyrics when you need exact wildcard behavior; those improvement tools may rewrite the text and overwrite or remove wildcard expressions.
Batch folder processing uses the same Wildcards behavior, so batch jobs can vary instruments, moods, hooks, or lyric phrases across outputs without manually editing every run.
For repeatable comparisons, save the manifest/settings and lock other variables such as model, duration, and seed while testing wildcard choices.

Write a concise Style prompt that describes genre, vocal character, instrumentation, production quality, tempo or mood, and mix target.
Write Lyrics with section tags such as [Verse] and [Chorus]. The included ACE_Step_Lyric_Generation_Instructions_For_LLMs.txt file can be given to an LLM to format lyrics or style prompts.
Optionally add Wildcards to Style or Lyrics when you want the app to choose between prompt variants automatically.
Select the Model. Start with ACE-Step XL 1.5 Turbo to verify the machine and workflow quickly.
Leave GPU Optimization Preset and DiT Quantization at safe defaults unless you are solving VRAM pressure or repeating a known workflow.
Set Song Duration and Songs. The demo run used 20 seconds and 1 song.
Use Random Seed while exploring. When a promising result appears, uncheck Random Seed and keep the seed so future edits stay comparable.
Click Generate Song and monitor the Status field plus the terminal window.

Demo generation result

Useful quick-tab buttons:

Random Style creates a starting style prompt.
Enhance Style improves the style prompt.
Enhance Lyrics improves lyric structure or phrasing before generation.
Cancel Generation stops a run from the UI.
Open Outputs Folder opens the app’s outputs directory where audio, manifests, lyrics, captions, sessions, and metadata are saved.

4. Results, Seeds, And Reuse

The tutorial stresses generating repeatedly until you have a good base result, then locking the seed and making controlled edits. This is especially important for remix and repaint work, where small prompt or range changes can be tested against the same underlying random state.

Seed and remix discussion

Results after generation

Send To Remix loads the generated song as Source Audio and prepares the advanced Remix workflow.
Send To Repaint loads the generated song and prepares a repaint range workflow.
Convert To Codes reuses the musical plan as LM Codes Hints in compatible Custom workflows.
Get Score and Get LRC create quality-score and lyric-timestamp artifacts.
Save and All Generated Files expose generated files for download or reuse.

Seed workflow:

Keep Random Seed on while searching for a usable base result.
When the result is close, copy or keep the seed shown by the UI.
Turn Random Seed off.
Change one word, one range, or one strength setting at a time.
Compare outputs against the locked seed.

5. Advanced Generation Modes

The ACESTEP Advanced tab is the full workstation. It exposes generation mode, runtime settings, source/reference audio, LM code utilities, advanced prompts, Wildcards in Music Caption/Lyrics, metadata, sampler settings, output settings, and batch processing.

Advanced overview

Generation modes:

Simple: plain-language generation when you want the app to fill many details.
Custom: precise manual control over caption, lyrics, BPM, key, time signature, language, duration, and advanced settings.
Remix: create a new version from source audio. The video recommends SFT for Remix and keeping the same lyrics for best results.
Repaint: replace or modify a selected time range while preserving the rest of the source audio.
Extract: use ACE-Step extraction or stem workflows where the selected model supports it.
LEGO: add a new stem to existing audio, such as adding a guitar stem to a track.
Complete: continue or complete source audio when the selected model supports it.

Advanced source audio

Advanced generation controls

Important advanced controls:

Source Audio is required for Remix, Repaint, Extract, LEGO, and Complete.
Reference Audio guides timbre, mix, performance feel, and atmosphere; it is not meant to copy exact melody, rhythm, or lyrics.
Analyze can fill or update caption/lyrics/metadata from source audio.
BPM Auto, Key Auto, TimeSig Auto, Language Auto, and Duration Auto let the model infer metadata.
Think enables LM planning. Turn it off only when deliberately using pasted LM Codes Hints.
Wildcards in Music Caption and Lyrics are expanded at generation time and also work when the same prompt fields are used for batch folder processing.
Retake creates controlled variation from the same seed/settings.
Edit changes the whole uploaded source using source and target prompts.
Auto Score, AutoGen, and Auto LRC can create score, metadata, and lyric timing outputs during generation.

Engine settings

Engine settings include GPU tier, checkpoint file, main model path, device, VAE, 5Hz LM model/backend, Flash Attention, CPU offload, compile, DiT quantization, LoRA path/folder, LoRA scale, inference steps, sampler, DCW, ADG, MP3 bitrate/sample rate, normalization, fades, LM temperature, top-k/top-p, negative prompt, and LM code settings. Leave these at defaults until you have verified a basic generation.

6. Remix, Repaint, Extract, LEGO, And Auto-Editor Features

The first part of the video demonstrates feature outcomes before the installation section. These are not separate apps; they are modes and panels inside the same ACE-Step interface.

Remix demo

Extract and LEGO demo

Auto-Editor demo

For Remix, upload the source song, keep the same lyrics, use SFT when available, and start with the default remix strength before changing one variable at a time.
For Repaint, set Repainting Start and End carefully, preview the selected range, then choose repaint mode and strength.
For Extract, choose Track Name or Extract All Stems. Batch folder processing can extract from multiple files.
For LEGO, choose or describe the new stem to add and use source audio as the foundation.
For Auto-Editor, set threshold, margin, mincut, and minclip. Workflow export can produce an editor timeline instead of rendering media.

7. Audio Processing

Audio Processing is used on uploaded or local audio/video and can also be applied automatically to generated songs. It includes format output, Auto-Editor trimming, video re-encode controls, audio enhancement stages, pre-mastering stages, DiffPitcher, and batch folder processing.

Audio Processing overview

Generated song loaded for processing

Audio Processing result

Core Audio Processing controls:

Apply automatically to generated songs runs the processing chain after generation.
Save original plus processed song keeps an untouched copy beside the processed copy.
Processed Output selects WAV/MP3 or another output format.
Processing Preset sets a preset chain before you tune individual stages.
Run as subprocess isolates processing so cancellation and memory cleanup are safer.
Export Only Audio extracts processed audio from video inputs.
Auto-Editor trim silent sections removes quiet/silent segments using threshold, margin, mincut, and minclip.
Auto-Editor workflow export exports an editing timeline/workflow instead of only rendering media.
Disable upload preview helps with very large MKV or multi-GB media.

Audio Enhancement and Pre-Mastering

Audio Enhancement includes Stereo Depth, Stereo Width, HF Refinement, Harmonic Enrichment, Timing Humanizer, and Ambience Shaping. Pre-Mastering includes Multiband Compressor, Tape Saturation, Glue Compressor, Mid/Side EQ, Soft Clipper, and LUFS Normalization.

DiffPitcher controls

DiffPitcher is for isolated vocals that sing the wrong notes. Use a guide vocal or MIDI score for the same phrase. The tutorial text in the UI warns that this is not for copying another singer or another song.

8. SAM Audio Segment

SAM Audio Segment is a heavier but more flexible segmentation system. It can extract target audio from a prompt, save the residual/remaining audio, process video inputs, use explicit span anchors, and run batch prompt lists separated by semicolons.

SAM Audio source-video demo

SAM Audio overview

SAM prompt runtime controls

Upload an audio or video file. Optionally upload a visual mask video for video-guided workflows.
Choose Mode and Quick Prompt, or type a Custom Prompt such as vocals, guitar, bass, drums, applause, or another target.
Enable Batch Segment when you want several prompts in one run; separate prompts with semicolons.
Use Predict spans when you want SAM Audio to estimate target time ranges from text.
Use explicit span anchor only when you can provide positive/negative time anchors as JSON.
Choose a VRAM preset and candidate count that match the GPU. Higher candidate counts can improve quality but cost runtime and VRAM.
Enable Save remaining audio when you need both the extracted target and the residual track.

9. Library, Metadata, Presets, Dataset, And Training Pages

The remaining app tabs are operational pages. They help you find previous generations, restore metadata, manage presets, inspect datasets, and train adapters.

Library

Load Metadata

Custom Preset System

Dataset browser

LoRA Dataset Builder

Train LoRA

Use Library when you want to find a past song by day and inspect its lyrics/metadata.
Use Load Metadata when you have a generation_manifest.json and want to restore a generation into the UI.
Use Custom Preset System to persist frequently used model, GPU, LoRA, audio, and generation defaults across sessions.
Use Dataset Builder to scan audio, auto-label captions/lyrics/BPM/key/time signature, review samples, and save a dataset JSON.
Use Preprocess before training; it creates tensor files for faster LoRA or LoKr training.
Use Train LoRA or Train LoKr only after preparing a clean dataset. The video states LoRA training is intended for a separate deeper tutorial.

10. RunPod Deployment

The RunPod chapter focuses on persistent network storage, GPU/region selection, unreliable installs, Gradio live URLs, nvitop monitoring, output downloads, and safe termination.

RunPod storage

RunPod install

RunPod Gradio services

RunPod monitoring

Create persistent network storage in the same region as the GPU you intend to rent.
Deploy the pod/template with the storage mounted. Choose a GPU with enough VRAM for the selected model and quality target.
Run the installer. If RunPod throws an OS/server error, run the installer again; it should resume from completed work.
If installation stalls from excessive parallelism, delete the virtual environment, lower installer thread count as shown in the video, and rerun.
Start the app and prefer the Gradio live link when the RunPod proxy is unreliable. If port 7860 does not open, try the port shown by the terminal, sometimes 7861.
Use nvitop to monitor GPU memory and load. First model load can be slow on RunPod storage; later generations are faster.
Download outputs from JupyterLab by right-clicking the outputs folder and downloading it as an archive.
Stop or terminate the pod deliberately. Delete storage too if you no longer want monthly storage charges.

11. Massed Compute Deployment

The Massed Compute chapter is similar to the Linux/cloud workflow, but the tutorial emphasizes faster disk performance and lower friction compared with RunPod. The tradeoff called out in the video is the lack of the same persistent network storage flow.

Massed Compute GPU selection

Massed Compute install

Choose the creator category and the SECourses image when following the video workflow.
Select a GPU appropriate for ACE-Step XL 1.5. The tutorial mentions RTX Pro 6000 and RTX 5090 class GPUs.
Upload the ACE-Step ZIP to Downloads, extract it, open Massed_Compute_Instructions_READ.txt, and copy the install command.
Open a terminal inside the extracted ACE-Step folder and run the command from that location.
Start ACE-Step and use the Gradio live URL. If Gradio live shows a transient error, refresh the page.
Back up large outputs or model/data folders to Hugging Face, Google Drive, OneDrive, or another storage service if you need to recreate the machine later.

12. SimplePod Deployment

The SimplePod chapter uses the RunPod/SimplePod instruction file and shows a persistent-storage flow that resembles RunPod. The tutorial demonstrates starting, generating, monitoring, stopping, and resuming from the same storage volume.

SimplePod setup

SimplePod generation

SimplePod resume

Register, add credits, and create/use persistent storage as shown in the instruction file.
Open the template link, attach the storage volume, choose a GPU, and run the machine.
Use the JupyterLab or console link to run the installer/start commands from the workspace.
If the Gradio live page throws a first-click error, refresh or click again after the page is fully loaded.
Install nvitop when you want GPU/VRAM visibility: pip install nvitop, then run nvitop.
To resume, reuse the template link, attach the same volume, select a GPU, start the machine, and run the app start command again.
Stop or terminate compute and remove storage when finished to avoid unwanted billing.

13. Troubleshooting And Best Practices

If Gradio errors after opening, refresh the browser or click again after the page finishes loading.
If RunPod installer errors, rerun the installer. If the virtual environment is corrupt, delete only the virtual environment and run the installer again.
If first generation is slow, wait for model load and monitor the terminal or nvitop.
For Remix quality, use SFT when available, keep lyrics aligned with the source, generate until you get a good base, then lock the seed.
For very large videos, use Disable upload preview in Audio Processing or supply a local path when available.
For prompt variation in many runs, use Wildcards in Style, Music Caption, or Lyrics. Batch folder processing can use the same wildcard syntax to vary each batch output. Keep Auto/Enhance Style and Auto/Enhance Lyrics disabled if you want exact wildcard expressions preserved.
Save presets and keep generation_manifest.json with outputs.
For VRAM pressure, use GPU Optimization Preset, quantization, offload controls, smaller duration/count, lower SAM candidates, or a larger GPU.
Stop/terminate cloud compute and delete storage volumes when the tutorial work is finished.

ACESTEP XL 1.5 Remix Mode Full Tutorial

Furkan Gözükara — Sun, 28 Jun 2026 00:57:50 +0000

ACESTEP XL 1.5 Remix Mode Full Tutorial

The ULTIMATE Local AI Update Just Dropped! ACE-Step 1.5, Paints-Undo & Whisper Premium Will WOW You

Full tutorial link > https://www.youtube.com/watch?v=hzKSt5WUAm0

Info

In this tutorial I show the newer ACE-Step XL 1.5 Premium features, especially the corrected remix workflow that was missing from the previous video. You will see how to remix songs properly, convert lyrics and language, tune remix strength and melody retention, regenerate only selected parts, use the Library metadata system, install the upgraded Paints-Undo pipeline, reduce VRAM usage, and fix repeating lines in Whisper Premium transcriptions.

Important links:

Download ACESTEP XL Premium files: [ https://www.patreon.com/posts/ace-step-1-5-xl-premium-157675060 ]
Discord: https://discord.com/invite/software-engineering-courses-secourses-772774097734074388
Google AI Studio: https://aistudio.google.com/
Previous ACE-Step tutorial: [ https://youtu.be/9C_6qNKjgpA ]
Windows requirements/setup tutorial: [ https://youtu.be/DrhUHnYfwC0 ]
Whisper Premium tutorial: [ https://youtu.be/4lAk6sf1qF8 ]
Download Whisper Premium App Files: [ https://www.patreon.com/posts/whisper-webui-premium-145395299 ]

Video chapters:

00:00 ACE-Step XL 1.5 Premium update, remix focus, and Paints-Undo preview
00:49 Billie Jean remix example: checking style preservation and vocal change
01:16 Lower remix strength idea and Gangnam Style Korean-to-English demo
01:36 Hearing the Korean-to-English result and why extreme remixes sound strange
01:53 Stronger remix settings with higher strength and melody retention values
02:12 How to update ACE-Step XL 1.5: download ZIP, extract, and overwrite
02:37 Run Windows install/update.bat and rebuild the virtual environment if needed
02:56 Library tab overview: daily categories, saved generations, and metadata
03:18 Loading old songs from Library with lyrics, parameters, and JSON restored
03:33 Starting a proper remix: select the SFT model and open Advanced -) Remix
03:45 Uploading the source song and learning the two key remix parameters
04:04 Remix presets explained: different lyrics, same lyrics, medium and big change
04:19 Why lower melody retention lets the model generate completely new lyrics
04:30 Torch compile speed tip and preparing the target style caption and lyrics
04:42 Using Gemini in Google AI Studio with the ACE-Step lyric instruction file
04:59 Editing or writing lyrics and matching the vocal language accurately
05:15 Launching a live remix generation and measuring local generation speed
05:27 Live timing result: around 33 seconds for a complete remix generation
05:49 Use generated result as source to repair or improve selected song parts
06:03 Selecting remix start/stop points and regenerating only the chosen section
06:17 How section patching works: full remix generated, only selected part replaced
06:36 When to keep lyrics/style the same and when to change them for a section
06:47 Comparing the full output against the newly generated section preview
06:58 Iterative remix workflow for perfecting each part of the composition
07:14 LoRA training progress, future voice accuracy, and language-swap limits
07:27 Fast local iterations, multiple attempts, no watermark, and usable outputs
07:55 Final remix reminders: SFT model, proper Windows setup, and compile mode
08:11 Tuning dramatic changes with percentage values and fixed seed comparisons
08:41 Paints-Undo upgraded intro: new pipeline, faster speed, and better results
08:52 Download, install, and start Paints-Undo with windows_startup.bat
09:08 First launch model downloads plus new xFormers Triton attention support
09:21 GPU compatibility, Torch 2.12.1, CUDA 13, and upgrades over the original
09:34 Upload an image, generate the prompt, and tag it with the WD14 tagger
09:52 Operating steps, keyframes, Tiled VAE options, and low VRAM preparation
10:04 24GB vs 7GB VRAM usage and how the new memory-saving options help
10:27 How keyframes become the drawing video before final video generation
10:52 CUDA 13, Torch 2.12, Triton attention, diagnostics, and speed improvements
11:10 Automatic attention fallback plus Linux/cloud installer compatibility notes
11:27 Full generation time, possible torch compile addition, and result preview
11:46 Why keyframes matter, why results vary, and why not every image works
11:57 Fixing out-of-VRAM errors by lowering resolution and using supported ratios
12:14 Whisper Premium update: transcribing videos with all Whisper model options
12:31 Faster Whisper quality mode and the repeated sentence problem
12:42 Using large-v1 and repetition penalty to prevent repeated subtitle lines
12:54 How to tune repetition penalty carefully so transcription is not skipped
13:06 Example result: highly accurate subtitles generated from the new video
13:19 29-minute video transcribed in 1.5 minutes, 20x real-time, and closing

Conclusion

This video is for users who want fast local AI music remixing, better generation iteration, image-to-drawing animation, and high quality subtitle transcription. Follow the timestamps to jump directly to ACE-Step remix settings, Paints-Undo installation, low VRAM options, or Whisper repetition penalty tuning.

App Installer Zip File Content

Some App Screenshots

Remix Page

Iterative Remix

Full Page

DEV Community: Furkan Gözükara

Install ComfyUI and Learn FLUX 3 Klein: The Complete Beginner Tutorial

Install ComfyUI and Learn FLUX 3 Klein: The Complete Beginner Tutorial

Contents

Links

Info

Chapters

18 October 2026 Update V143

NEW PRESETS — MOSS-Image: Text To Image, Fast, Turbo, Layout Boxes and Base

UPDATED PRESET — AvatarForever: Live Preview Panel and Tested Mouth-Movement Settings

NEW PROMPT GUIDE — MOSS-Image Feed for ChatGPT, Claude and Gemini

NEW — ComfyUI Colab Master Notebook for Google Colab

LAUNCHER AND INSTALLER FIXES

HOW TO UPDATE

17 October 2026 Update V142

NEW PRESETS — FLUX 3 Klein: Text To Image, Image Edit and Layout Boxes

NEW MODEL — Our FLUX 3 Klein HQ INT8 ConvRot

NEW PROMPT GUIDE — FLUX 3 Klein Feed for ChatGPT, Claude and Gemini

AvatarForever Preset Now Has Live Preview Feature

HOW TO UPDATE

10 October 2026 Update V140

New MiniMax H3 Presets

MiniMax H3 Long Lip Synch - Native Latent Continuation - Full Audio

MiniMax H3 RefMod Builder - Images, Video, Audio

MiniMax H3 RefMods: Reusable References

MiniMax H3 Speed Optimizer: Native Sparse Attention

Other MiniMax H3 Improvements

Works With the Latest ComfyUI

Preset Fixes and Updates

Installer Improvements

Zero to AI-Ready PC in One Click: Windows & Ubuntu AI Requirements Installer Full Tutorial

Zero to AI-Ready PC in One Click: Windows & Ubuntu AI Requirements Installer Full Tutorial

Links:

Chapters:

1-Click Windows and Linux (Ubuntu 22, 24, 26) AI Requirements Installers - Setup Your PC and Make It Ready for All Local AI Apps, Scripts, and More

Why Our AI Requirements Installers

What You Download

Windows AI Requirements Installer

How to Run It

What It Installs

20 Steps, Every One Timed

The Latest NVIDIA Driver

Made for PCs That Already Have Tools

PATH in the Right Order

Environment Variables

Visual Studio Updates Keep Compiling Working

78 Tests, Run for Real

Checked Downloads and Backups

Windows Options

Ubuntu AI Requirements Installer

How to Run It

What It Installs

17 Steps, the NVIDIA Driver Last

Ready in Every New Terminal

66 Tests, Run as You

Before and After, Downloads and Backups

Ubuntu Options

Run It Again Any Time

Better Than ElevenLabs? Ultimate TTS Voice Training & Cloning in 600+ Languages

Links:

Video Chapters:

Why Ultimate Text To Speech Generator

1-Click Installation

What You Download

Latest PyTorch and CUDA 13

Automatic Model Downloads

Start the App

Cloud GPUs: RunPod, SimplePod and Massed Compute

Requirements

Three Speech Models in One App

GPU Presets for 6 GB to 32 GB Cards

Voice Cloning From Any Audio or Video

Text, Smart Sentences and Pauses

Pronunciation Check and Dictionary

Take Quality: The Best Take of Every Section

OmniVoice: Clone, Design or Auto Voice

AuK: Describe a Voice in Words

Emotion Control

Timing, Output Formats and Word Timestamps

Subtitles to Timed Speech and MP4