MOSS-Image, OpenMOSS’s 5B open-weights text-to-image model, released on October 16, 2026, under Apache-2.0, now has native support in ComfyUI. This isn’t a custom node pack situation; the support was merged via pull request #17095 on the same day the model dropped. For builders, this means the integration is clean, but the documentation highlights two critical settings that templates get wrong by default. If you want Chinese and English text rendering or JSON-based layout control, you need to know exactly which files go where and which knobs to turn.

Version Check and File Placement

Before downloading gigabytes of weights, verify your ComfyUI version. The standard text-to-image templates require ComfyUI 0.42.0 or later, while the newer Image Edit and Color & Depth templates need 0.43.0+. A quick check involves opening a Load CLIP node; if 'moss_image' doesn't appear in the type dropdown, your install is outdated. The weights are hosted on the Comfy-Org/MOSS-Image repository on Hugging Face. You need four specific files: the SFT checkpoint (moss_image_sft_int8_convrot.safetensors) in models/diffusion_models/, the Qwen3.5 9B text encoder (qwen3.5_9b_w4a8_gptq.safetensors) in models/text_encoders/, and the FLUX.2 small VAE (flux2-vae_small.safetensors) in models/vae/. Note that the text encoder is larger (7.34 GB) than the main checkpoint (5.41 GB), which is crucial for VRAM planning.

Tuning for Quality Over Speed

The default templates are optimized for speed, resulting in undercooked images for final output. OpenMOSS recommends increasing the steps from the default 25 to 50 or more and raising the CFG scale from 2.5 to 3 (or 3-6 for short prompts). You should also switch the sampler from er_sde to Euler (flow matching) and ensure canvas dimensions are multiples of 16, targeting 1 to 2 megapixels. Two settings you can leave alone: shift is built-in at 5, and skip layer guidance is active at layer 22. Disabling SLG might boost realism but risks collapsing fine details.

Key Takeaways

  • Native support means no custom nodes or pip installs are required for basic generation.
  • The Qwen3.5 text encoder is the largest file (7.34 GB), exceeding the model checkpoint size.
  • Default settings prioritize speed; bump steps to 50+ and CFG to 3 for usable results.
  • ComfyUI’s native scheduler won't reproduce Diffusers images exactly due to grid-snapping differences.

The Bottom Line

This is a solid native integration, but don't trust the default template settings. Adjust the steps and CFG immediately, or you’ll waste time debugging what is actually just a low-quality default render.

Next Steps for Builders

If you need precise layout control, look into the box-drawing nodes in the layout templates, which allow you to assign captions to specific regions. For training, stick to the SFT checkpoint for inference and the Base checkpoint for training. If you’re experimenting with LoRA, tools like KohakuMossTrainer and ai-toolkit are supported, but remember to keep the text encoder frozen and cache encoded captions to save time.