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

Original Patreon post

Full tutorial: Watch on YouTube

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.

Contents

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

Links

• ComfyUI package
• Model downloader
• Windows requirements tutorial
• Requirements installer
• Cloud setup tutorials

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

1. Download the latest ComfyUI_V143.zip from this post.
2. Extract it over your existing installer folder and overwrite the older files.
3. 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.
4. 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.
5. Restart ComfyUI and refresh the browser tab.
6. 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

1. Download the latest ComfyUI_V141.zip from this post.
2. Extract it over your existing installer folder and overwrite the older files.
3. 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.
4. 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.
5. Restart ComfyUI and refresh the browser tab.
6. 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.