Local setup
MOSS Image ComfyUI guide
MOSS Image runs natively in ComfyUI 0.42.0 and later. Pick your GPU memory and this MOSS Image ComfyUI guide shows which files to download, where they go and the settings to start with.
Your GPU memory
Under 8 GB: run it online
Even the smallest community build plus the text encoder needs more memory than these cards have, so ComfyUI keeps swapping to system RAM and images come out slowly. Skip the download and run MOSS Image online, with 6 free credits for a new account.
Run MOSS Image online8 GB: community GGUF build
Load the Q4_K_S GGUF through the ComfyUI-MOSS-GGUF and ComfyUI-GGUF custom nodes and start ComfyUI with --lowvram. This route is community made and experimental; its author tested it on an RTX 3070.
Download
- MOSS-Image-SFT-Q4_K_S.gguf3.46 GBPut it in ComfyUI/models/diffusion_models/
- MOSS-Qwen3.5-9B-INT8-ConvRot.safetensors8.45 GBPut it in ComfyUI/models/text_encoders/
- MOSS-VAE.safetensors0.25 GBPut it in ComfyUI/models/vae/
Total download: 12.2 GB
Start with
- 1024 x 1024, 50 steps, CFG 3, euler sampler, sgm_uniform scheduler
- Shift 5, skip layer guidance on block 22, empty negative prompt
- Launch ComfyUI with --lowvram
12 GB: the official int8 set
These are the files ComfyUI's own MOSS-Image templates load. A community test on a 12 GB RTX 4070 ran a larger set, with the int8 text encoder, at 832 x 1216 in about 51 seconds for 50 steps.
Download
- moss_image_sft_int8_convrot.safetensors5.41 GBPut it in ComfyUI/models/diffusion_models/
- qwen3.5_9b_w4a8_gptq.safetensors7.34 GBPut it in ComfyUI/models/text_encoders/
- flux2-vae_small.safetensors0.25 GBPut it in ComfyUI/models/vae/
Total download: 13.0 GB
Start with
- Template defaults: 25 steps, CFG 2.5, er_sde sampler
- For more detail: the official 50 steps and CFG 3
- 1 to 2 megapixels, width and height in multiples of 16
16 GB: int8 model, int8 text encoder
The same int8 model with a less compressed text encoder and the full VAE, and room for 2 megapixel images.
Download
- moss_image_sft_int8_convrot.safetensors5.41 GBPut it in ComfyUI/models/diffusion_models/
- qwen3.5_9b_int8_convrot.safetensors9.75 GBPut it in ComfyUI/models/text_encoders/
- flux2-vae.safetensors0.34 GBPut it in ComfyUI/models/vae/
Total download: 15.5 GB
Start with
- Official values: 50 steps, CFG 3
- Up to 2 megapixels, such as 1152 x 1728 for a poster
24 GB and up: the full bf16 model
The model at full precision, as OpenMOSS released it, with the int8 text encoder. With 32 GB or more you can swap in the 19.3 GB bf16 text encoder too.
Download
- moss_image_sft_bf16.safetensors10.16 GBPut it in ComfyUI/models/diffusion_models/
- qwen3.5_9b_int8_convrot.safetensors9.75 GBPut it in ComfyUI/models/text_encoders/
- flux2-vae.safetensors0.34 GBPut it in ComfyUI/models/vae/
Total download: 20.3 GB
Start with
- Official values: 50 steps, CFG 3
- Up to 4 megapixels (2048 x 2048)
Mac: no tested route yet
ComfyUI's MOSS-Image guide has no Mac instructions. A community Turbo int8 build (8 steps, CFG 1) was made for testing on macOS, but nobody has published Apple Silicon speeds. To skip the experiment, run it online.
Run MOSS Image onlineFile sizes are from the Hugging Face repos (Comfy-Org/MOSS-Image, realrebelai/MOSS-Image_GGUFs), checked October 20, 2026. Memory tiers come from community tests and file sizes, not our own runs; real use depends on resolution and what else is on the GPU. Sources are listed at the end of this page.
Files
Where the MOSS Image ComfyUI files go
Three files make a working MOSS Image ComfyUI setup: the diffusion model, the Qwen3.5 9B text encoder and the FLUX.2 VAE. Each goes in its own models folder.
- UNET Loader reads the diffusion model
- CLIPLoader reads the text encoder, with type set to moss_image
- VAE Loader reads flux2-vae_small or flux2-vae
Full list of builds: Comfy-Org/MOSS-Image on Hugging Face.
# 12 GB set, from Comfy-Org/MOSS-Image on Hugging Face$ mv moss_image_sft_int8_convrot.safetensors ComfyUI/models/diffusion_models/$ mv qwen3.5_9b_w4a8_gptq.safetensors ComfyUI/models/text_encoders/$ mv flux2-vae_small.safetensors ComfyUI/models/vae/# Restart ComfyUI so the loaders list the new filesSetup
How to set up MOSS Image in ComfyUI
From an existing ComfyUI install to the first image in five steps.
- 01
Update ComfyUI to 0.42.0 or later
The 0.42.0 release on October 18, 2026 lists MOSS-Image support in its notes; ComfyUI's own guide names 0.41.3 as the minimum. The desktop app updates itself; a git install needs a pull, the portable build its update script.
- 02
Download the files for your GPU
Pick your memory in the guide at the top of this page. The 12 GB set is the one the official templates expect.
- 03
Move them into the models folders
diffusion_models, text_encoders and vae, then restart ComfyUI.
- 04
Open a MOSS-Image template
Browse Templates and search "MOSS-Image". There are four: Text to Image and Layout-Guided Text to Image, each for SFT and Base. Start with SFT.
- 05
Describe the finished image and run
Write what the picture looks like, put every word to print in quotes and add line breaks where the text should wrap.
Settings
MOSS Image ComfyUI settings: template vs official
The ComfyUI templates trade some detail for speed. When an image looks soft or unfinished, switch to the values OpenMOSS publishes.
| Feature | ComfyUI template | OpenMOSS recommendation |
|---|---|---|
| Steps | 25 | 50 |
| CFG | 2.5 | 3 |
| Sampler | er_sde | Euler (flow matching) |
| Shift | In the template subgraph | 5 |
| Skip layer guidance | On, layer 22 | On, layer 22 |
| Negative prompt | nsfw, worst quality, bad anatomy, watermark, jpeg artifact | Empty |
| Resolution | 1 megapixel | 1 to 2 megapixels |
Speed
Faster MOSS Image in ComfyUI
Fifty steps on a 5B model take a while on consumer cards. Three ways to cut the wait, from simplest to most involved.
Keep the template at 25 steps
Half the official step count, already set in the official templates. Raise it only for images that need finer detail.
Turbo int8 build: 8 steps
A community 8-step version of the SFT model (toxicdog/MOSS-Image-SFT-Turbo-INT8-ConvRot, 5.41 GB) that drops into UNET Loader. Run it at 8 steps, CFG 1, euler, simple. Experimental: compare it with the standard model on your own prompts.
Step-distilled LoRAs: for tinkerers
There are 4-step and 8-step LoRAs (akhaliq, darask0, lodestones), but they ship extra time embeddings and some need an fp32 model and their own sampler. Wait for a ready-made workflow unless you enjoy wiring nodes.
Troubleshooting
Fix common MOSS Image ComfyUI errors
The MOSS Image ComfyUI problems people hit most in the first week, and what fixes each one.
- 01
Missing nodes, or no moss_image type in CLIPLoader
Your ComfyUI is older than 0.42.0. Update it, or install the OpenMOSS MOSS-Image plugin from ComfyUI Manager.
- 02
Every image is pure noise
A ModelSamplingSD3 node in the workflow causes it. Replace it with ModelSamplingAuraFlow or ModelSamplingFlux, or start again from the official template, which has none (ComfyUI issue 17119).
- 03
Out of memory
Switch to the int8 files, stay at 1 megapixel, start ComfyUI with --lowvram, or move to the GGUF route on 8 GB cards.
- 04
Text comes out garbled
Quote every word to print, add real line breaks and keep each block short. For exact placement, use the Layout-Guided template.
- 05
A GGUF file will not load
GGUF builds need the ComfyUI-MOSS-GGUF and ComfyUI-GGUF custom nodes and their own loaders. The W4A8 safetensors model files do not load in that node.
- 06
Images take far too long
Your GPU is short of memory and ComfyUI is swapping to system RAM. Use a smaller set from the guide at the top, or run MOSS Image online.
Sources
Checked October 20, 2026. Memory tiers and speeds come from these community reports, not our own runs.
- ComfyUI MOSS-Image guide: files, folders, template settings, minimum version 0.41.3
- ComfyUI 0.42.0 release notes: MOSS-Image support
- Comfy-Org/MOSS-Image on Hugging Face: official file sizes
- ComfyUI issue 17119: pure noise from ModelSamplingSD3
- RTX 4070 test: int8 model and int8 text encoder on 12 GB, about 51 seconds per 50-step image
- realrebelai/MOSS-Image_GGUFs: GGUF builds, tested on an RTX 3070 in low-VRAM mode
- GGUF Versions discussion: GGUF is not supported in the official integration
- toxicdog/MOSS-Image-SFT-Turbo-INT8-ConvRot: 8-step Turbo build, made for macOS testing
Skip the setup and run MOSS Image online
No download and no GPU. Type a prompt or plan a layout in the browser; new accounts start with 6 free credits.
FAQ
MOSS Image ComfyUI questions
Yes. Native MOSS-Image support is listed in the ComfyUI 0.42.0 release notes (October 18, 2026), and ComfyUI's own guide asks for 0.41.3 or later. Official templates cover SFT and Base. Only the community GGUF builds need custom nodes.
A 12 GB card runs the official int8 set at 1 megapixel. 8 GB cards need the community GGUF build and --lowvram, and the full bf16 model is most comfortable on 24 GB. These figures come from community tests (for example this RTX 4070 run) and file sizes, so leave some headroom.
In ComfyUI, open Browse Templates and search "MOSS-Image". The same four MOSS Image ComfyUI workflows are JSON files in the Comfy-Org workflow_templates repository on GitHub, for example image_moss_image_t2i.json and image_moss_image_t2i_bboxes.json for layouts.
There is no tested route yet. The official guide has no Mac instructions, and the community Turbo int8 build made for macOS has no published speeds. MOSS Image online runs in any browser.
No. A Comfy-Org member said GGUF is not supported in the official integration; the GGUF builds come from the community (realrebelai/MOSS-Image_GGUFs) and load through their own custom nodes.
It takes a box for each headline and object, so every word lands where you drew it. The MOSS Image layout editor on our home page writes the same JSON layout without coordinates by hand.