A refined and enhanced workflow by Experimental_dAIver.

This workflow brings the strong prompt adherence, layout control and text rendering of MOSS-Image SFT (the 5B DiT from OpenMOSS) into your local ComfyUI setup. It is tuned for the Abiray INT8 ConvRot quant and features dedicated custom loaders for the diffusion model, text encoder and VAE, NdSuperLoraLoader for seamless multi-LoRA stacking with automatic trigger-word handling, AILab_Resolution_Selector for clean latent sizing, the official Uncond-skip guidance, a complete integrated SEEDVR2 Video Upscaler Subgraph and professional output via CiviScribe.

Version 1.0 is the initial public release and delivers a clean, flexible base specifically tuned for the MOSS-Image SFT INT8 checkpoint.

Key features in v1.0:

• Full native support for the moss-image-sft-dit-int8-convrot diffusion model via the dedicated MOSSImageUNETLoader (from the OpenMOSS custom node pack)

• Dedicated MOSSImageCLIPLoader for the Qwen3.5-based text encoder and standard VAELoader for the matching VAE

• NdSuperLoraLoader (nd-super-nodes) with full multi-LoRA support, auto-fetch trigger words and clean strength control

• AILab_Resolution_Selector – convenient aspect-ratio and megapixel selector (different node from my earlier workflows) with automatic multiple-of-16 calculation

• SkipLayerGuidanceDiTSimple (“Uncond skip (empty = off)”) for the official layer-skip guidance (single_layers = 22 / 24 / 26)

• Advanced sampling via KSampler (er_sde or euler) + ModelSamplingSD3 (shift 5.0) with sgm_uniform or bong_tangent scheduler

• Complete SEEDVR2 Video Upscaler Subgraph – the usual high-quality DiT-based upscaler for 4K+ stills and short sequences, with Lab color correction and intelligent resolution handling

• Professional output with CiviScribe SaveImage (CCollins_CiviScribe_SaveImage). The official node works perfectly; I personally use a lightly modified version for my own CivitAI uploads

• Self-documenting workflow containing a detailed Note with download links, recommended settings (with and without the speed LoRA) and folder guidance

• Integrated MOSS-Image Prompt Input subgraph that cleanly concatenates trigger words with your positive prompt

Required Custom Nodes

OpenMOSS MOSS-Image custom node – MOSSImageUNETLoader, MOSSImageCLIPLoader
https://github.com/OpenMOSS/MOSS-Image
Copy the comfyui/moss-image folder into ComfyUI/custom_nodes/moss-image (or install via ComfyUI Manager once indexed).

nd-super-nodes (HenkDz) – NdSuperLoraLoader
https://github.com/HenkDz/nd-super-nodes

comfyui-rmbg – AILab_Resolution_Selector
(Install via ComfyUI Manager)

CiviScribe (CCollins) – CCollins_CiviScribe_SaveImage
https://github.com/ccollins/CiviScribe
(The official version is sufficient. I use a personal modification for my own metadata workflow.)

ComfyUI-SeedVR2_VideoUpscaler – full SEEDVR2 Video Upscaler subgraph
https://github.com/ainvfx/ComfyUI-SeedVR2_VideoUpscaler

Important: Update ComfyUI to a recent version before installing the custom nodes. The OpenMOSS loaders and native MOSS-Image support benefit from current core updates.

Models & Downloads (exact paths)

The following list explains exactly which files you need and where to place them. Names match the Note embedded in the workflow.

1. Main Diffusion Model (MOSS-Image SFT INT8 ConvRot):

File: moss-image-sft-dit-int8-convrot.safetensors
Target folder: ComfyUI/models/diffusion_models/Moss/
(Create the Moss subfolder if it does not exist. The MOSSImageUNETLoader references this path.)

2. Text Encoder:

File: moss-image-sft-qwen35-int8-convrot.safetensors
Target folder: ComfyUI/models/text_encoders/

3. VAE:

File: moss-image-sft-vae.safetensors
Target folder: ComfyUI/models/vae/

4. Optional Speed LoRA (DMD):

File: MOSS_Image_SFT_8_steps_dmd_v0.1.safetensors (or current equivalent)
Target folder: ComfyUI/models/loras/Moss/
Use only for fast drafts (see special note below).

5. SEEDVR2 Models (for the integrated upscaler subgraph – optional but highly recommended):

DiT Model: seedvr2_ema_7b_fp16.safetensors (or your preferred quant)
VAE: ema_vae_fp16.safetensors

Download from the official ComfyUI-SeedVR2_VideoUpscaler repository or Hugging Face and follow the node’s installation instructions.

Folder Structure Overview

📂 ComfyUI/
├── 📂 models/
│   ├── 📂 diffusion_models/
│   │   └── 📂 Moss/
│   │       └── moss-image-sft-dit-int8-convrot.safetensors
│   ├── 📂 text_encoders/
│   │   └── moss-image-sft-qwen35-int8-convrot.safetensors
│   ├── 📂 vae/
│   │   └── moss-image-sft-vae.safetensors
│   └── 📂 loras/
│       └── 📂 Moss/
│           └── (optional DMD speed LoRA)

Recommended Settings

Without speed LoRA (quality mode):

Sampler: er_sde (recommended) or euler
Scheduler: sgm_uniform or bong_tangent
Steps: 30–50
CFG Scale: 3.0
Shift (ModelSamplingSD3): 5.0
Uncond skip: single_layers = 22 (24 or 26 also useful)
Resolution: multiples of 16 (1 MP recommended via AILab_Resolution_Selector)

With speed LoRA (fast drafts):

Sampler: er_sde
Scheduler: sgm_uniform or bong_tangent
Steps: 8–10
CFG Scale: 1.0
Shift: 5.0
Uncond skip: must be bypassed (see note below)
Resolution: same as above

A compact settings table is also embedded directly in the workflow’s Note node for quick reference.

How to use the workflow

1. Update ComfyUI and install the required custom nodes listed above.

2. Download the models and place them exactly in the folders specified.

3. Load the JSON workflow file into ComfyUI.

4. Select your desired aspect ratio and megapixels in the AILab_Resolution_Selector.

5. Enter your prompt in the Positive Prompt node (the NdSuperLoraLoader automatically handles trigger words). The model understands both natural language and structured JSON layout prompts.

6. Load one or more compatible LoRAs into the NdSuperLoraLoader if desired.

7. Generate. The workflow produces a high-quality base image.

8. (Optional) Run the integrated SEEDVR2 upscaler subgraph for beautiful high-resolution results.

9. Save with the CiviScribe node (includes metadata and optional CivitAI manifest).

Special Notes

Speed LoRA (DMD) for fast drafts
The DMD speed LoRA is excellent when you only need quick previews. It works reliably at 8–10 steps with CFG 1.0.
Critical: When the speed LoRA is enabled you must bypass the “Uncond skip (empty = off)” node. Leaving it active produces data salad. With the node bypassed the 8–10 step runs are clean and usable for drafting.

NSFW behaviour
The model is fully uncensored. It will generate explicit content without additional tools or LoRAs.
However, once the motif becomes NSFW the results turn very wonky. The more hardcore the scene, the more it becomes a pure dice roll whether you receive a clean, coherent image or collapsed anatomy and artefacts. SFW and mild scenes are considerably more reliable.

This is a clean, practical base workflow for the MOSS-Image SFT INT8 quant. You can mute or bypass the upscaler group entirely if you only need the base generation.

Huge thanks to the OpenMOSS team for the model and custom nodes, to the CiviScribe author, and to the SeedVR2 maintainers.

If you have questions, feedback or want to share your results – I’m always happy to hear from you in the comments. Enjoy the workflow and happy creating!