MOSS-Image ComfyUI Workflow: Models, Templates and Settings

MOSS-Image is the open-weights text-to-image model OpenMOSS released on October 16, 2026: a 5B single-stream diffusion transformer (DiT) under the Apache-2.0 license, best known for rendering Chinese and English text and for following JSON layouts. ComfyUI supports it natively, so you do not need a custom node pack.

This guide covers the four official ComfyUI templates, the model files and where they go, and the settings worth changing. If you just want to see what the model can do first, our MOSS Image AI generator runs MOSS-Image-SFT in the browser with the official 50 steps and CFG 3.

What you needCopy Anchor Link

• ComfyUI 0.42.0 or later. The MOSS-Image templates need it. Native support was merged into ComfyUI on October 16 (pull request #17095). The official repository notes that the v0.41.3 release did not include it, so an older install will report missing nodes.
• A GPU with room for the int8 checkpoints, which the templates load by default. bf16 builds are in the same repository if you have more VRAM.
• About 13 GB of disk for the files the SFT templates use, or about 18.4 GB with the Base checkpoint too.

The four official templatesCopy Anchor Link

In ComfyUI, open Template and search for MOSS-Image. Four templates ship, and all four run the same subgraph, so you can open it to reach the loaders, the sampler and the canvas size.

You can also download each template as JSON from Comfy-Org/workflow_templates and drag it onto the canvas.

Not sure which checkpoint you want? Read MOSS-Image Base vs SFT. The short version: SFT for making images, Base for training.

Download the model filesCopy Anchor Link

All ComfyUI-format files are in Comfy-Org/MOSS-Image on Hugging Face. The templates load these four:

You only need the Base checkpoint if you plan to use the Base templates.

ComfyUI/
└── models/
    ├── diffusion_models/
    │   ├── moss_image_sft_int8_convrot.safetensors
    │   └── moss_image_base_int8_convrot.safetensors
    ├── text_encoders/
    │   └── qwen3.5_9b_w4a8_gptq.safetensors
    └── vae/
        └── flux2-vae_small.safetensors

The same repository also has:

• bf16 builds of both checkpoints, 10.16 GB each: moss_image_sft_bf16.safetensors and moss_image_base_bf16.safetensors. If you have the VRAM, swap the checkpoint in the diffusion model loader (UNET Loader).
• Larger builds of the text encoder (int8 at 9.75 GB, bf16 at 19.31 GB) and the full flux2-vae.safetensors.
• A color and depth control LoRA, moss_image_color_depth_control.safetensors, for models/loras/. It is repackaged from darask0/MOSS-Image-color-depth-control.

Run your first imageCopy Anchor Link

1. Update ComfyUI to 0.42.0 or later. If you cannot find the MOSS-Image templates in the library, your install is too old.
2. Go to Template, search for MOSS-Image and open MOSS-Image: Text to Image.
3. Put the four files above in their folders, then refresh or restart ComfyUI so the loaders can see them.
4. Type your idea into the positive prompt and click Queue. The caption the model actually received appears in the Preview Any node.

Because the text-to-image template runs your prompt through the upsampler, a one-line idea works. This prompt is already a full caption, and it is the one behind the image at the top of this page:

Prompt

A wide conference banner with the title "AI 开发者大会 2026" and the subtitle "AI Developer Summit". A clean blue and violet gradient background, abstract circuit lines, and a bottom row reading "10月28日 · 深圳 · Oct 28 · Shenzhen".

Try this prompt

Settings that matterCopy Anchor Link

The templates are tuned for speed. OpenMOSS's own recommendations are slower, and they are what the official examples use:

A few notes on these:

• If a result looks undercooked, switch to 50 steps at CFG 3 before you change anything else. The ComfyUI docs give the same advice.
• Skip layer guidance (SLG) is on by default at layer 22, the layer OpenMOSS recommends. The subgraph node has an SLG toggle. The official user guide says that turning it off makes images look very realistic but can make some details collapse.
• Canvas size. MOSS-Image is trained for 1 to 2 megapixels and extrapolates from about 0.25 MP (512×512) to 4 MP. Width and height must be multiples of 16. The ResolutionSelector node picks the aspect ratio and the megapixel count for you.
• Same seed, different image. ComfyUI's native scheduler does not use the training-grid snapping of the official Python pipeline, so the same seed and settings will not reproduce a Diffusers image exactly. That is expected.

The prompt upsamplerCopy Anchor Link

The text-to-image templates include a prompt upsampler that rewrites a short request into a full scene caption before sampling. It runs on the Qwen3.5 text encoder the model already loads, so there is nothing extra to download.

When you want full control of the wording, set enable_prompt_upsampler to false on the subgraph node. Your prompt then goes to the model unchanged. Either way, Preview Any shows the caption that was used, so you can compare.

Writing prompts by hand? The MOSS-Image prompt guide covers how to get clean Chinese and English text, line breaks included.

Layout-guided generationCopy Anchor Link

The two layout templates add a pair of nodes:

1. Create Bounding Boxes: drag one box per element and type a caption for each box. For a text box, type the text you want rendered.
2. MOSS-Image Layout Prompt: turns your boxes into the layout prompt the model reads. Your overall scene prompt is used as well.

Boxes are the most reliable way to put a title, a product shot or a block of text exactly where you want it. Describing positions in prose works less well. MOSS-Image reads layouts as JSON with a bounding box per element; if you want to write or generate that JSON yourself, see the MOSS-Image JSON layout tutorial.

Native support vs the repository pluginCopy Anchor Link

Before native support landed, the official repository shipped a ComfyUI plugin with a MOSS-Image HF Loader node that loads a local Hugging Face / Diffusers directory. It is still there, but the repository now recommends the native path: the built-in Load Diffusion Model, CLIP Loader (type moss_image) and Load VAE nodes, with no plugin. Some older pages, including the model cards on Hugging Face, still tell you to install a plugin from ComfyUI Manager. The repository's ComfyUI README is newer.

If you are moving an old plugin workflow over:

• Import a native workflow instead of swapping only the model loader. The loader nodes are different.
• Do not carry over the plugin workflow's ModelSamplingSD3 node unchanged: the native model uses a different timestep multiplier. To change the shift of the native model, use ModelSamplingAuraFlow with multiplier 1.

TroubleshootingCopy Anchor Link

• Missing nodes, or no MOSS-Image templates. Update ComfyUI. The templates need 0.42.0 or later, and a brand-new core node can reach the nightly build before a stable release. ComfyUI Cloud updates after stable releases.
• Undercooked results. Go to 50 steps and CFG 3. With a very short prompt, leave the upsampler on or raise CFG.
• Vertical color banding with SFT. Community LoRAs exist to reduce it. The official README lists MOSS-Image-SFT-DeBand by MATLOWAI, which comes with ComfyUI preset workflows. Its model card explains the trade-off between less banding and changes to fine detail.
• Running in FP16 instead of BF16. MOSS-Image-FP16-fixed by KBlueLeaf, also listed in the official README, provides FP16-compatible weights for Base and SFT in Diffusers and ComfyUI formats.
• Too slow. ComfyUI-Spectrum-MOSS accelerates sampling, and there is a community few-step distillation (see the next section).

Is there a MOSS-Image Turbo?Copy Anchor Link

Not from OpenMOSS, as of October 25, 2026. The official family is MOSS-Image-Base and MOSS-Image-SFT, and the user guide says few-step distillation methods such as DMD are still in the works. The "turbo" checkpoints and LoRAs you may come across are community projects. One that the official README lists is lodestones' MOSS-Image-SFT-8-steps-dmd, an 8-step DMD distillation of SFT with LoRA adapters for ComfyUI. Follow each model card's recommended settings, since few-step models usually need different steps and CFG.

Training a LoRACopy Anchor Link

Train on Base. It has no post-training or CFG distillation, which OpenMOSS says is meant to keep the full training signal and make LoRA training easier. Their advice is to keep the text encoder frozen. Two trainers listed in the official README support both checkpoints:

• KohakuMossTrainer: full fine-tuning and LyCORIS (including LoRA), with adapter training on as little as 6 GB of VRAM, on Windows and Linux.
• DiffSynth-Studio: full fine-tuning and LoRA training.

Try MOSS-Image without installing anythingCopy Anchor Link

No GPU, or want to test a prompt before you download 13 GB of files? Our free MOSS-Image generator runs the SFT checkpoint online. Click the button below to load this prompt into it:

Prompt

A rainy evening street scene in front of a small noodle shop with a glowing neon sign reading "老李面馆" and the English line "Lao Li Noodles" below it. A hand-written menu board next to the door lists "牛肉面 Beef Noodles 18" and "葱油拌面 Scallion Noodles 15". Warm window light, wet pavement reflections.

Try this prompt

SourcesCopy Anchor Link

• OpenMOSS/MOSS-Image on GitHub and its ComfyUI README
• MOSS-Image user guide
• Comfy-Org/MOSS-Image on Hugging Face
• MOSS-Image in ComfyUI, ComfyUI documentation