How do I run MOSS-Image in ComfyUI?
Last updated 2026-10-27
Needs ComfyUI 0.42.0+ (edit needs past v0.42.1). Open Workflow→Browse Templates→MOSS-Image. You need 3 files: a diffusion model, the Qwen3.5-9B text encoder (type moss_image) and the FLUX.2 small VAE.
No GPU? Try MOSS-Image text posters in the editor →
Which files do I need?
| File | Size | Folder |
|---|---|---|
moss_image_sft_int8_convrot / bf16 | 5.41 / 10.16 GB | models/diffusion_models |
moss_image_base_int8_convrot / bf16 (for training LoRAs) | 5.41 / 10.16 GB | models/diffusion_models |
qwen3.5_9b_w4a8_gptq / int8_convrot / bf16 | 7.34 / 9.75 / 19.31 GB | models/text_encoders |
flux2-vae_small | 0.25 GB | models/vae |
moss_image_edit_lora_rank_64 / color_depth_control | 0.30 / 0.31 GB | models/loras |
Download from Comfy-Org/MOSS-Image. The ComfyUI MOSS-Image tutorial lists the same files.
How do I set it up?
- Update ComfyUI to 0.42.0 or newer (past v0.42.1 for edit).
- Download the diffusion model, text encoder and VAE into the folders above.
- Open Workflow → Browse Templates → MOSS-Image and pick a template.
- In the Load CLIP node, set the type to
moss_image. - Run with the settings below.
Which model and encoder fit my VRAM?
| Memory | Diffusion model | Text encoder |
|---|---|---|
| Low VRAM | int8 (5.41 GB) | w4a8 (7.34 GB) |
| Mid | int8 (5.41 GB) | int8 (9.75 GB) |
| High VRAM | bf16 (10.16 GB) | bf16 (19.31 GB) |
These pairs are combinations of the listed files, not measured VRAM requirements. Community builds for low memory exist: GGUF and SDNQ 4-bit versions from OzzyGT, and a 4-step GGUF from mischiefaudio.
Which official templates are there?
- Text-to-image (SFT)
- Text-to-image (Base)
- Layout-guided (SFT)
- Layout-guided (Base)
- MOSS-Image Edit, with the edit LoRA. See MOSS-Image Edit.
What settings should I use?
| Setting | Value |
|---|---|
| Steps | 50 |
| CFG | 3 |
| Shift | 5 |
| Sampler | Euler |
| Skip-layer guidance | Layer 22 |
| 8-step DMD LoRA | 8 steps, CFG 1 |
Prompt syntax, the negative prompt and layout JSON are in the prompt guide.
What about images above 2 MP and the prompt enhancer?
The ComfyUI tutorial has sections on resolutions above 2 MP and on the prompt enhancer. Read them in the official docs.
How do I fix common errors?
- Node not found: update ComfyUI.
- Text encoder will not load: set the Load CLIP type to
moss_image. - Out of memory: switch to the int8 diffusion model and the w4a8 text encoder.
Where is the Kromoss fine-tune?
Kromoss is a community fine-tune of MOSS-Image-Base by lodestones. It uses the same text encoder and VAE as MOSS-Image-SFT. The OpenMOSS repository has the code and user guide.
FAQ
How much VRAM do I need?
The sources list file sizes, not VRAM. The bf16 diffusion model and text encoder total about 29.5 GB on disk; the int8 diffusion model plus the w4a8 text encoder total about 12.8 GB. For low VRAM use int8 and w4a8. Community GGUF and SDNQ 4-bit builds also exist.
Should I use Base or SFT?
Use SFT to generate images; it lays out text more cleanly. Use Base when you want to train a LoRA.
Can I use MOSS-Image with diffusers instead of ComfyUI?
A community diffusers version exists: OzzyGT/MOSS_Image_SFT_diffusers on Hugging Face, published 2026-10-27.
Why does ComfyUI say a node is missing?
Your ComfyUI is too old. MOSS-Image needs 0.42.0 or newer, and the edit nodes need a version past v0.42.1. Update ComfyUI and reopen the template.
What does the 8-step LoRA change?
The 8-step DMD LoRA replaces 50 steps with 8 and uses CFG 1 instead of 3.