1. Start with native MOSS Image support
The recommended MOSS Image ComfyUI workflow uses native loaders. OpenMOSS documents support from ComfyUI commit e087d8f6, merged October 16, 2026. Use a current build containing that change and its matching dependencies.
Restart ComfyUI after updating. You should be able to select moss_image in the CLIP loader. The official installation guide notes that the older v0.41.3 release does not contain this support.
Use native single-file models for the main steps below. The optional HF-directory plugin has different loaders and scheduling; its separate setup appears later in this guide.
You will need local storage for the diffusion model, text encoder and VAE, plus hardware capable of running your chosen precision. Model file size alone is not a GPU memory requirement. Quantized weights are available from Comfy-Org.
2. Download and place the model files
Open Comfy-Org’s MOSS Image files. The native SFT template checked for this guide uses the following files:
| File | Folder inside ComfyUI | Loader |
|---|---|---|
moss_image_sft_int8_convrot.safetensors | models/diffusion_models/ | Load Diffusion Model |
qwen3.5_9b_w4a8_gptq.safetensors | models/text_encoders/ | CLIP Loader, type moss_image |
flux2-vae_small.safetensors | models/vae/ | Load VAE |
The same repository also provides BF16 alternatives. If you select an alternative file, update its loader selection too. For Base, choose the Base checkpoint and Base workflow together. Read the Base vs SFT explanation if you are deciding between them.
3. Import the native ComfyUI workflow
Save the appropriate upstream JSON file, then drag it into ComfyUI or open it with the workflow loader. These files come from Comfy-Org’s workflow repository.
- Open the generation subgraph and confirm all three model loader selections match the files you installed.
- Enter a subject, scene and composition. Put any words you want in the picture in quotation marks.
- For a simple first run, turn off the template’s prompt enhancer and enter the text prompt directly. A reference-image input in the template is for its prompt enhancer, rather than an image-editing input to MOSS.
- Start with a 1024 × 1024 canvas and the imported template’s sampling defaults. Run the workflow and inspect the saved image.
- Change one setting at a time. Keep the seed fixed when comparing prompts, and check small lettering before publishing.
| Steps | CFG | Sampler | Scheduler |
|---|---|---|---|
| 25 | 2.5 | er_sde | simple |
These are the imported native template’s settings, not a universal preset. The optional plugin example and Kromoss author’s recommendations differ. Keep each workflow’s model, guidance and scheduling configuration together.
4. Use Kromoss in a compatible ComfyUI workflow
Kromoss is lodestones’ community fine-tune of MOSS-Image-Base. Its model card provides compatible single-file weights and the author’s sampling recommendations.
- Place
kromoss_bf16.safetensorsorkromoss_int8_convrot.safetensorsinmodels/diffusion_models/. - Select it in Load Diffusion Model, retaining the MOSS-compatible text encoder and VAE.
- For the standard Kromoss run, use the author’s 50-step, CFG 3 configuration.
Optional 8-step Kromoss LoRA
Put one supplied DMD or DMD2 LoRA in models/loras/. The author recommends DMD2. Connect LoraLoaderModelOnly after the diffusion model loader at strength 1, then use Euler, the simple scheduler, 8 steps and CFG 1. Disable negative-branch layer skipping for this path.
Use one acceleration LoRA at a time. This is a local Kromoss configuration; our online generator continues to run MOSS-Image-SFT.
5. Optional: load a Hugging Face model directory
If you need the repository’s HF-directory loader, follow the separate OpenMOSS plugin instructions. Copy comfyui/moss-image/ from that repository into ComfyUI/custom_nodes/moss-image/, then restart.
Import the plugin workflow JSON. Download the complete Base model directory and point MOSS-Image HF Loader to the local folder containing transformer/, text_encoder/, tokenizer/ and vae/. A Hub repository name alone is not a local directory.
The plugin example uses 50 steps, CFG 3, shift 5, Euler / sgm_uniform and negative-branch skipping at layer 22. Do not copy its ModelSamplingSD3 node unchanged into a native workflow: native scheduling differs. OpenMOSS identifies ModelSamplingAuraFlow as the native shift adjustment node when needed.
MOSS Image ComfyUI troubleshooting
The moss_image loader type is missing
Check the ComfyUI build, update its dependencies and restart. Import a native workflow for native loaders; installing the optional plugin is a separate route.
A model file is missing from a loader
Check its folder, exact filename and completed download. Refresh or restart ComfyUI, then reselect the file. A Diffusers directory and a single-file checkpoint are different formats.
Generation runs out of memory
Use the compatible quantized files, reduce the canvas and close other GPU workloads. The encoder and VAE also need memory, so a diffusion checkpoint’s download size is not the total working memory.
A Kromoss LoRA produces poor output
Check the model compatibility and LoRA strength. For the documented 8-step route, use CFG 1 with Euler / simple and no negative-branch skipping. Do not retain the standard run’s CFG 3 settings.
I only want to try MOSS Image
Use the online text-to-image generator without local installation, or browse prompt ideas. The online model is SFT; Kromoss is covered here as a local workflow option.