comfyui-moss-Image
Native MOSS-Image loaders for ComfyUI that output standard MODEL, CLIP and VAE objects, so you can use ComfyUI's own sampling, guidance and memory management nodes.
MOSS-Image for ComfyUI
English | 中文
Native MOSS-Image loaders for ComfyUI that output standard MODEL, CLIP and VAE objects, so you can use ComfyUI's own sampling, guidance and memory management nodes.
Forked from OpenMOSS/MOSS-Image (the comfyui/ subtree), republished as a standalone custom node.
Installation
Option 1: Comfy CLI / ComfyUI-Manager (recommended)
comfy node install comfyui-moss-Image
Or search for comfyui-moss-Image in ComfyUI-Manager → Custom Nodes → Install.
Option 2: Manual git clone
cd ComfyUI/custom_nodes
git clone https://github.com/wanaigc/comfyui-moss-Image.git
Restart ComfyUI. The plugin uses ComfyUI's bundled Python dependencies, no extra install step is required.
Model Files
| Component | File | Size | Format |
| ---------------- | ---------------------------------------------------------- | ------- | ----------------------------------------------- |
| DiT | diffusion_models/moss_image_sft_int8_convrot.safetensors | 5.25 GB | INT8 ConvRot (Modulation, Layer 0 & 41 in BF16) |
| Text Encoder | text_encoders/qwen35_moss_int8_convrot.safetensors | 8.65 GB | INT8 ConvRot (SSM & Layer 31 in BF16) |
| VAE | vae/flux2-vae.safetensors | 238 MB | FP32 (FLUX.2 Small Decoder, 384ch) |
Comfy-Org official quantized build: Quark Pan · HuggingFace · ModelScope
XCEED123 INT8 ConvRot build: Quark Pan (国内加速) · HuggingFace
Original weights (OpenMOSS): MOSS-Image-Base · MOSS-Image-SFT
Place the files in your ComfyUI directories:
📂 ComfyUI/
├── 📂 models/
│ ├── 📂 diffusion_models/
│ │ └── moss_image_sft_int8_convrot.safetensors
│ ├── 📂 text_encoders/
│ │ └── qwen35_moss_int8_convrot.safetensors
│ └── 📂 vae/
│ └── flux2-vae.safetensors
Recommended Settings
| Setting | Value | Notes |
| ----------------------- | ------------------------- | --------------------------------------------------------- |
| Sampler | euler | Flow-matching Euler |
| Scheduler | sgm_uniform | Pairs with flow-matching shift |
| Steps | 50 | 30–50 steps |
| CFG | 3.0 | 2.5–3.5 |
| Shift | 5.0 | Set in ModelSamplingSD3 node |
| Skip Layer Guidance | 22 | In SkipLayerGuidanceDiTSimple, set single_layers="22" |
| Resolution | 1024×1024 or widescreen | Multiples of 16 |
Documentation
See the upstream installation and usage guide.
The example workflow lives at example_workflows/image_moss_sft_int8_t2i.json.
License
This project is available under the Apache-2.0 license, same as the upstream OpenMOSS/MOSS-Image repository.