Guide
How do I use MOSS-Image in ComfyUI?
Use a ComfyUI nightly build from October 16, 2026 or later, download the MOSS-Image files from Comfy-Org on Hugging Face into three model folders, and load the official “MOSS-Image: Text to Image” workflow. ComfyUI added native MOSS-Image support on October 16, 2026 (commit e087d8f6). As of October 19, the stable release v0.41.3 did not include it yet.
What you need before you start
- An NVIDIA GPU with CUDA. Official support is CUDA only; see the VRAM guide for memory.
- ComfyUI master/nightly at commit
e087d8f6or later, with its currentrequirements.txtinstalled. - About 16–30 GB of disk space, depending on whether you choose the int8 or bf16 files.
Step 1 – Download the files
From the Comfy-Org model repository, pick one file from each group:
| Folder | Pick one | Size |
|---|---|---|
models/diffusion_models/ |
moss_image_sft_bf16.safetensors (best quality) or moss_image_sft_int8_convrot.safetensors (smaller) |
10.2 GB / 5.4 GB |
models/text_encoders/ |
qwen3.5_9b_bf16.safetensors, qwen3.5_9b_int8_convrot.safetensors or qwen3.5_9b_w4a8_gptq.safetensors |
19.3 / 9.8 / 7.3 GB |
models/vae/ |
flux2-vae.safetensors or flux2-vae_small.safetensors |
0.34 / 0.25 GB |
Start with SFT for everyday generation; consider Base for training or fine-tuning. See the version comparison.
Step 2 – Load the workflow
Open the Comfy-Org workflow templates and import MOSS-Image: Text to Image. For position control, use MOSS-Image: Layout-Guided Text to Image. The workflow uses Load Diffusion Model, CLIP Loader set to type moss_image, and Load VAE.
Step 3 – Use the recommended settings
50 steps, CFG 3, shift 5, Euler sampler, at 1024×1024. Native nodes already default to shift 5. To change it, use ModelSamplingAuraFlow with multiplier 1. Text encoding is capped at 4,096 tokens.
Common problems and fixes
- “Unknown model type” or missing nodes: your ComfyUI is too old. Update to a nightly build that includes commit
e087d8f6. - Old plugin workflow gives strange results: the native model uses a different timestep multiplier. Import a native workflow instead of swapping only the loader, and don’t reuse the plugin’s ModelSamplingSD3 node.
- Out of memory: switch to the int8 diffusion model and the w4a8 text encoder, or lower the resolution. See the VRAM guide.
- Same seed, different image from the Python pipeline: expected. The official README notes that ComfyUI’s scheduler differs, so matching seeds and settings don’t guarantee identical images.
Don’t want to set it up?
Open MOSS-Image AI to generate with the SFT model online, without installing ComfyUI. Eligible visitors can try one Standard image without signing up, subject to network and daily limits. See the free-use guide.
Questions
Do I still need the MOSS-Image custom node plugin?
No. Native nodes are recommended. The plugin is only for loading a local Hugging Face folder or for old plugin workflows.
Which text encoder should I pick?
bf16 for best fidelity if memory allows; int8 or w4a8 to save memory. All three are in the Comfy-Org repo.
Is there a layout workflow?
Yes. "MOSS-Image: Layout-Guided Text to Image" includes the MOSS-Image Layout Prompt node.
Can I use LoRAs?
Yes. LoRAs with matching keys work with the built-in LoRA loader.
Does it work on the stable ComfyUI release?
Not as of October 19, 2026. You need a nightly build from October 16 or later.
Turn your idea into an image
Describe your idea, generate an image, then refine the text or layout if needed.