DEV Community

AI-KSK
AI-KSK

Posted on

Six MOSS-Image ComfyUI Workflows in One Tutorial: Character Reference, QHD, SFT Style, Portraits and Low-Step Generation

#ai #comfyui #machinelearning #tutorial

Download all six original workflow JSON files and covers.

I explored a range of community workflows around MOSS-Image in a single AI-KSK video tutorial. Rather than treating every preset as a new model, this article walks through six separate ComfyUI graphs from that episode: one for character references, one for direct QHD rendering, two for style and portrait LoRA experiments, one Kromoss eight-step pipeline, and one three-branch sampling comparison.

You can watch the full episode on YouTube or Bilibili. The six JSON files are collected together for this episode; no third-party model weights are redistributed in the workflow package.

The six modules at a glance

Module Workflow JSON Task Exported resolution and steps
01 01_MOSS_CharRef_角色参考生图.json Character reference conditioning 832×1216, 50 steps
02 02_MOSS_MAX_ULTRA_QHD_QHD高清生图.json Direct widescreen QHD generation 2560×1440, 40 steps
04 04_MOSS_SFT_Style_Lab_SFT风格实验.json Kijai rank-256 SFT-difference LoRA experiment 1024×1024, 25 steps
05 05_MOSS_Beauty_人像美感创作.json Beauty portrait LoRA 1024×1024, 30 steps
06 06_MOSS_Kromoss_8Step_8步文生图.json Kromoss DMD2 low-step generation 1024×1024, 8 steps
07 07_MOSS_StepLab_50_8_4_50·8·4步三档生图.json Three full/accelerated SFT recipes 1024×1024, 50 / 8 / 4 steps

These IDs follow the original file numbering. There is no module 03 among the six supplied exports. Every graph is available separately inside the episode's repository folders; the six accompanying 3:4 posters are promotional cover designs, not validated generation samples.

Shared setup: do this once

These graphs depend on native MOSS-Image support in ComfyUI. The OpenMOSS ComfyUI guide documents the supported build and loader scheme. Native support merged on October 16, 2026; an older stable ComfyUI build may lack the corresponding nodes. Use a build that includes MOSS support and install the requirements appropriate to that build before importing the graphs.

The common loader files are in Comfy-Org/MOSS-Image:

ComfyUI/models/
  diffusion_models/
    moss_image_base_int8_convrot.safetensors  # modules 01, 02
    moss_image_sft_int8_convrot.safetensors   # modules 04, 05, 07
    kromoss_int8_convrot.safetensors          # module 06; different upstream source
  text_encoders/
    qwen3.5_9b_bf16.safetensors
  vae/
    flux2-vae.safetensors
  loras/
    [each module-specific adapter named in its section below]
Enter fullscreen mode Exit fullscreen mode

The Kromoss backbone comes from lodestones/Kromoss, not from the Comfy-Org MOSS files. Import a JSON, let ComfyUI report any absent dependencies, and select the exact files named in the graph. The text-encoder loader uses the moss_image type. Do not assume every community LoRA is interchangeable with either Base, SFT or Kromoss.

Each workflow is a starting preset, not a GPU compatibility certificate. The exported JSON contains sampler values, prompts and some historical notes, but it does not include reproducible performance telemetry. The video's subjective comparisons are discussed as creator observations, not universal benchmark results.

1. Character reference: MOSS CharRef v2

File: 01_MOSS_CharRef_角色参考生图.json. Promotional cover: 01_MOSS_CharRef_封面_3x4.png.

This module is for supplying an image as Picture 1 and asking for a new scene while referring to the character. It connects MOSS-Image Base INT8, 7err4's CharRef v2 LoRA, and custom nodes from ostris/ComfyUI-MOSS-Image-Ostris-Edit. It is not the same as training an identity LoRA or guaranteeing exact identity preservation.

Active settings: CharRef strength 1.0, 832×1216 latent, 50 steps, CFG 3, Euler / sgm_uniform, seed 42, and skip-layer guidance at layer 22. Additional nodes such as TextEncodeMOSSImageOstrisEdit and MOSSImageOstrisEditModelPatch must be available.

To use it, first install the Ostris extension and the CharRef adapter, alongside Base, Qwen3.5 and the VAE. Open the JSON, replace the LoadImage node's file with a photo or illustration you have permission to use, and rewrite the prompt so its clothing and identity cues actually match that reference. Confirm the reference enters the edit-conditioning path, run a trial and check the output against your own input. The archive does not include the long-hash PNG referenced by the exported LoadImage node. The poster is not a substitute for that missing reference.

Identity consistency depends on the reference, subject, prompt and model limitations. No identity preservation score, before/after example, or guaranteed face fidelity is supplied here.

2. MAX ULTRA QHD: direct 2560×1440 generation

File: 02_MOSS_MAX_ULTRA_QHD_QHD高清生图.json. Cover: 02_MOSS_MAX_ULTRA_QHD_封面_3x4.png.

This graph generates a new image from text using MOSS-Image Base INT8 and comfyanonymous's MAX ULTRA high-resolution adapter. It initializes an empty latent at 2560×1440, rather than taking a low-resolution image and upscaling it. The selected LoRA is moss_hires_lora_test_v0.4.safetensors at 1.0. The sampler runs 40 steps, CFG 3, Euler / simple, seed 9527, with denoise 1.

Install the Base checkpoint and matching text encoder/VAE, add the named v0.4 hires adapter to models/loras/, import the graph and verify the latent dimensions and sampler. Execute a local test before drawing conclusions about quality or memory use. The note box inside the file mentions 2048×2048, 25 steps and a 38.4 GB memory observation, but those numbers are not the active resolution and step settings, and the memory number is not backed by independent telemetry in this package. Do not quote it as a tested requirement. Higher resolution can increase memory pressure; if you reduce the latent size, record that as a separate modified run.

3. SFT Style Lab: a rank-256 difference-LoRA experiment

File: 04_MOSS_SFT_Style_Lab_SFT风格实验.json. Cover: 04_MOSS_SFT_Style_Lab_封面_3x4.png.

This is a style-oriented exploration of Kijai's experimental MOSS-image adapters. The graph loads MOSS-Image SFT INT8 and then overlays moss_image_sft_lora_rank_256_bf16.safetensors via a model-only LoRA loader. The positive text describes a retro anime-style profile in a red, black and gold visual vocabulary. That is the prompt in the file, not proof of a particular final image.

The exported LoRA strength is 1.49; its embedded historical Note says 1.0. Use the active 1.49 when describing this unmodified export. The sampler uses 1024×1024, 25 steps, CFG 3, Euler / simple, seed 9527. In the video narration, I explain the idea of approximating a Base-to-SFT difference with an adapter and changing its strength. That design rationale does not create an independently replicated 1.0-versus-1.49 image comparison in this package.

Install the SFT checkpoint, Qwen3.5 encoder, FLUX.2 VAE and the exact rank-256 file, then import and run the JSON. To study strength effects responsibly, duplicate the graph outside the distributed original, vary only the LoRA strength in your new experiment and record the output images and environment. The upstream experimental LoRA repository does not clearly specify a license in its model-card metadata; use the upstream download link rather than rehosting weights.

4. Beauty Portraits: a portrait-oriented LoRA

File: 05_MOSS_Beauty_人像美感创作.json. Cover: 05_MOSS_Beauty_封面_3x4.png.

This graph attaches the axiomofmind MOSS Image Beauty LoRA to the SFT INT8 backbone. Use the explicitly named MOSS-Image-Beauty-LoRA-ComfyUI.safetensors: the upstream repository distinguishes this native ComfyUI-format file from a separate Diffusers/PEFT-root adapter. The file runs 1024×1024, 30 steps, CFG 3, Euler / simple, seed 42, with LoRA strength 1.0 and skip-layer guidance at 22.

Install the ComfyUI-format LoRA, open the graph and check the positive prompt. The upstream author recommends the trigger phrase moss_beauty_style, but that trigger is not present in the supplied active prompt. You can evaluate it in a separately saved variation, but do not describe it as part of this unmodified export. The supplied creator narration discusses its intended portrait aesthetic; no independent sharpness, realism or facial-quality score was measured for this package.

5. Kromoss: an eight-step pipeline with its own backbone

File: 06_MOSS_Kromoss_8Step_8步文生图.json. Cover: 06_MOSS_Kromoss_8Step_封面_3x4.png.

Do not replace Kromoss with stock SFT. This graph loads lodestones' Kromoss fine-tuned INT8 checkpoint kromoss_int8_convrot.safetensors and its matching 8-steps-dmd2-lora.safetensors, strength 1.0. The other shared files remain Qwen3.5 and FLUX.2 VAE. Active parameters: 1024×1024, eight sampler steps, CFG 1, Euler / simple, seed 9527.

Place both the Kromoss model and matching DMD2 adapter in the designated folders, import the JSON and verify their pairing. Try a prompt and compare the visual behavior with SFT if you want to investigate fine-tune differences. Eight steps refers to diffusion sampling count, not eight seconds, and the supplied package does not contain GPU time measurements or a validated speedup ratio.

6. Triple-Step Studio: three different recipes in one graph

File: 07_MOSS_StepLab_50_8_4_50·8·4步三档生图.json. Cover: 07_MOSS_StepLab_50_8_4_封面_3x4.png.

This graph contains three text-to-image branches with the same MOSS SFT INT8, shared positive prompt, 1024×1024 latent and seed 4242. It provides a practical visual comparison of full-sampling and distilled recipes; it is not a pure experiment changing only the step count.

Branch Required adapter/conditioning Steps CFG Sampler/scheduler Save prefix
Full Skip-layer guidance 22; no speed LoRA 50 3 Euler / simple M06_50step
DMD2 lodestones DMD2, 8-steps-dmd2-lora.safetensors at 1.0 8 1 Euler / simple M06_8step
Turbo akhaliq four-step Turbo, moss_image_sft_4step_v2.0.safetensors at 1.0 4 1 Euler / simple M06_4step

The negative-conditioning routes also differ, so this is a comparison of whole inference recipes. Install both adapters, import the graph and inspect the three individual SaveImage outputs. My supplied episode narration describes the 50-step example as looking better to me in that demonstration; that remains a subjective observation, not a statistically supported quality ranking. Record actual timings on your own environment if performance matters.

Which workflow should you choose?

If your immediate task is reference-conditioned scene creation, start with CharRef, but bring your own source image. For direct widescreen generation at the documented setting, choose MAX ULTRA QHD and assess your hardware. Choose SFT Style Lab for controlled experiments with the rank-256 difference adapter, or Beauty for a dedicated portrait treatment. Use Kromoss to explore a separately fine-tuned eight-step model. Use Triple-Step Studio when you want to inspect the tradeoffs of full, DMD2 and Turbo recipes while keeping their different CFG/adapters in view.

None of the posters shipped with the collection is claimed as a verified rendering from its paired JSON. The model weights have their own authors and licenses; this resource set does not redistribute them.

Episode and resources

Attribution: Workflow compilation and episode narration: AI-KSK; third-party models/adapters belong to the named upstream developers. Specific redistribution terms for the JSON exports and promotional covers must be confirmed by the asset owner. A license on one model does not grant blanket rights over the entire episode or other model components.

Try these workflows online and download individual graphs

Workflow RunningHub Civitai
01 · MOSS-Image CharRef v2 Character Reference RunningHub Workflow
02 · MOSS MAX ULTRA QHD Text-to-Image RunningHub Workflow
04 · MOSS-Image SFT Style Lab RunningHub Workflow
05 · MOSS-Image Beauty Portrait Workflow RunningHub Workflow
06 · Kromoss 8-Step Text-to-Image RunningHub Workflow
07 · MOSS-Image 50/8/4-Step Comparison Lab RunningHub Workflow

RunningHub links include the AI-KSK referral code. The hosted versions may change independently of these exported JSON files; check the live settings before running.

Top comments (0)