Characters, illustrations, and little objects — ready for your composition.

中文 · English · Download models · Model code

Generate RGBA images with MOSS-Image and save PNG files with an alpha channel, using ComfyUI's normal model loading, prompting, and sampling nodes. **The DiT LoRA is optional.** The original MOSS-Image model can generate images with this RGBA VAE alone. The optional DiT LoRA adapts MOSS-Image-SFT for transparent-image generation. Transparency and edge quality vary with the prompt and combination. ## Install 1. Update ComfyUI and prepare the base model and text encoder using the [official MOSS-Image instructions](https://github.com/OpenMOSS/MOSS-Image/blob/main/comfyui/README.md). 2. Clone this repository into `ComfyUI/custom_nodes/`: ```bash cd ComfyUI/custom_nodes git clone https://github.com/Chenkin-x/ComfyUI-CKMoss-Image-RGBA.git ``` 3. Restart ComfyUI. This package requires no additional Python dependencies. The entry point should be `ComfyUI/custom_nodes/ComfyUI-CKMoss-Image-RGBA/__init__.py`. ## Download weights Both files live in [one HF repository](https://huggingface.co/ChenkinLab/CKMoss-Image-RGBA/tree/main). | File | Destination | Purpose | | --- | --- | --- | | [moss_rgba_vae.safetensors](https://huggingface.co/ChenkinLab/CKMoss-Image-RGBA/resolve/main/vae/moss_rgba_vae.safetensors) | `ComfyUI/models/vae/` | Required; merged RGBA VAE | | [moss_rgba_dit_lora.safetensors](https://huggingface.co/ChenkinLab/CKMoss-Image-RGBA/resolve/main/loras/moss_rgba_dit_lora.safetensors) | `ComfyUI/models/loras/` | Optional LoRA for MOSS-Image-SFT | These are a VAE and an adapter. You also need the MOSS-Image base model and its text encoder. ## Generate Drag a workflow JSON into ComfyUI: - [RGBA VAE only](../workflows/vae_only.json): no DiT LoRA. - [RGBA VAE + DiT LoRA](../workflows/with_lora.json): uses `Load LoRA (Model Only)`, with strength `1.0`. Choose your MOSS-Image-SFT model in `Load Diffusion Model`, its text encoder in `Load CLIP`, and `moss_image` as the CLIP type. Choose this project's VAE in **Load MOSS RGBA VAE**. Describe the subject, appearance, pose, and style. You can append `transparent background`. The examples use Euler, 30 steps, CFG 4, and dimensions divisible by 16. **VAE Decode → Save Image** preserves the RGBA PNG. For separate color and mask processing, use **Split RGBA Image**. Its mask follows ComfyUI conventions: white is transparent, black is opaque. Use `Join Image with Alpha` to combine them again. Examples below were generated with the RGBA VAE and DiT LoRA. Display backgrounds and checkerboards are separate from the original PNG files. CKMoss Image RGBA examples

## Credits Built on [MOSS-Image](https://github.com/OpenMOSS/MOSS-Image) and the [FLUX.2 VAE](https://github.com/black-forest-labs/flux2#flux2-autoencoder). Thanks to [niangao2331](https://github.com/niangao2331) for contributing part of the image data. Code is licensed under GPLv3. See [NOTICE](../NOTICE) for upstream attribution.