SFT or Base.
MOSS-Image is OpenMOSS’s 5B single-stream diffusion transformer, released on 16 October 2026. It works in the Flux.2 latent space, reads prompts with Qwen3.5 9B and decodes with the Flux.2 VAE. It’s known for clear text in pictures, in any language, for layouts that keep things in place, and for being easy to fine-tune.
- MOSS-Image SFT is tuned on curated data. It’s the one to draw with.
- MOSS-Image Base is the pre-training checkpoint, for training LoRAs.
Both use the same encoder and VAE, and the same settings.
Files you need.
One model, one encoder, one VAE, all from Comfy-Org/MOSS-Image, which isn’t gated. Each model comes in bf16 and in int8_convrot, which is about half the size.
-
Model5.4 GB Download
moss_image_sft_int8_convrot.safetensorsComfyUI/models/diffusion_models/ or moss_image_sft_bf16.safetensors, 10.2 GB -
Base5.4 GB Download
moss_image_base_int8_convrot.safetensorsComfyUI/models/diffusion_models/ optional, for LoRA training. bf16 is 10.2 GB -
Text encoder9.7 GB Download
qwen3.5_9b_int8_convrot.safetensorsComfyUI/models/text_encoders/ or qwen3.5_9b_w4a8_gptq.safetensors, 7.3 GB -
VAE0.3 GB Download
flux2-vae.safetensorsComfyUI/models/vae/
The VAE is the same flux2-vae that Flux.2 uses, so if you already have it, you’re set. The encoder is the plain Qwen3.5 9B; FLUX.3’s prompt-upsampler file has the same shapes but is tuned differently, so don’t use it here.
models/
├── diffusion_models/
│ └── moss_image_sft_int8_convrot.safetensors
├── text_encoders/
│ └── qwen3.5_9b_int8_convrot.safetensors
└── vae/
└── flux2-vae.safetensors
What fits your computer.
At 1024 × 1024. The encoder runs first and ComfyUI moves it out of the way before sampling, so the model file sets the limit. Few people have posted numbers yet, so treat the middle tiers as estimates from file sizes.
- 8 GBTight
int8 model (5.4 GB). The encoder is moved in and out, so system RAM matters; 32 GB is comfortable. On an RTX 3070-class card, 50 steps with CFG take roughly 4 to 5 minutes, going by one user’s run.
- 12 GBFits
int8 comfortably. bf16 (10.2 GB) leaves little room.
- 16 GBFits
bf16, at the recommended 1 to 2 megapixels.
- Large GPUsFits
The official Diffusers folder is full precision, about 27 GB of files, and needs a large card.
- MacUntested
No Apple Silicon reports yet.
Set it up.
-
Update ComfyUI
ComfyUI’s own MOSS-Image support landed on 16 October and shipped in 0.42.0 on 18 October. Older versions can’t load the single files. ComfyUI Desktop updates itself; in a manual install:
Terminal, in the ComfyUI foldergit pull pip install -r requirements.txt
-
Download the three files
Model, encoder and VAE from Comfy-Org/MOSS-Image, using the list above. Take SFT unless you’re training a LoRA.
-
Put them in their folders
Model in
diffusion_models, encoder intext_encoders, VAE invae. Restart ComfyUI. -
Check the loaders
Load Diffusion Model gets the MOSS-Image file. Load CLIP gets
qwen3.5_9bwith typemoss_image, and the latent comes fromEmptyFlux2LatentImage. -
Set the steps and run
50 steps at CFG 3 with euler and sgm_uniform, as the makers sample. ComfyUI’s own nodes apply shift 5 by themselves, so there’s no shift node to add.
Settings that work.
These are the makers’ settings, from the model card and their workflow.
- Steps
- 50
- CFG
- 3
- Sampler
- euler
- Scheduler
- sgm_uniform
- Shift
- 5built in on ComfyUI’s nodes
- Size
- 1 to 2 MP0.25 to 4 MP work. Multiples of 16
- Negative
- yes
- Text encoder
- Load CLIP, moss_image
1024 × 1024 is a good start. Any size from 0.25 to 4 MP works, with both sides in multiples of 16, but the makers recommend staying between 1 and 2 MP.
Two ways to run it.
Comfy-Org’s single files with ComfyUI’s own nodes, as above, are the smaller download. OpenMOSS also ships its own way:
- The MOSS-Image node pack (Comfy Registry id
moss-image) from OpenMOSS works on ComfyUI 0.38 or newer. - It loads the official Diffusers folder, OpenMOSS-Team/MOSS-Image-SFT: about 27 GB in full precision, with the text encoder and VAE included. Put the whole folder in
ComfyUI/models/diffusers/. - It also skips layer 22 on the negative pass, as the makers sample.
When it goes wrong.
- The single files won’t load, or the text encoder type is unknown
- ComfyUI is older than 0.42.0. Update it, or use the MOSS-Image node pack with the Diffusers folder.
- Wrong text encoder picked up
- FLUX.3’s prompt-upsampler file has the same shapes as Qwen3.5 9B but isn’t MOSS-Image’s encoder. Use
qwen3.5_9b_int8_convrotorqwen3.5_9b_w4a8_gptq. - Out of memory with the Diffusers folder
- It’s full precision, about 27 GB. On smaller cards use the int8 single file instead.
- Very slow on an 8 GB card
- 50 steps with CFG at 1024 × 1024 take minutes there, and the encoder swaps through system RAM. 32 GB of system RAM is comfortable.
- Not sure the Base file is the right one
- Base is the pre-training checkpoint, meant for LoRA training. Draw with SFT.
Questions.
Should I download MOSS-Image SFT or Base?
SFT, to make pictures. It’s tuned on curated data. Base is the pre-training checkpoint and is meant for training LoRAs.
Which ComfyUI version does MOSS-Image need?
ComfyUI 0.42.0 or newer for its own nodes and Comfy-Org’s single files. OpenMOSS’s MOSS-Image node pack, which loads the official Diffusers folder, works from ComfyUI 0.38.
What settings does MOSS-Image use?
The makers’ settings are 50 steps, CFG 3, shift 5, the euler sampler and the sgm_uniform scheduler, at 1 to 2 megapixels with sides in multiples of 16.
Can I use FLUX.3’s prompt upsampler as the MOSS-Image text encoder?
No. It’s a tuned Qwen3.5 9B with the same shapes, but MOSS-Image reads prompts with the plain Qwen3.5 9B from Comfy-Org/MOSS-Image.
Can I use MOSS-Image commercially?
Yes. MOSS-Image is released under Apache 2.0.
Sources: OpenMOSS MOSS-Image repository, MOSS-Image-SFT model card, Comfy-Org MOSS-Image files, MOSS-Image node pack on the Comfy Registry.