This is a ComfyUI node for MatAnyone, a state-of-the-art video matting model that can remove backgrounds from videos using just a single mask for the first frame for enhanced/guided video matting.
MatAnyone only requires that you provide the first single frame alpha mask (solid white for the subject against a solid black background for the stuff you don't want in the final output). My example workflow uses the Rembg+ Session nodes from Comfy Essentials to automatically create the first frame alpha mask for you. The dance video example and the anime example videos below both used that method. Alternatively, you can provide the first frame alpha mask yourself and bypass those nodes. It's really that simple. Check the workflow directory in this repo for the example workflow.
- Clone this repository into your ComfyUI custom_nodes folder:
cd ComfyUI/custom_nodes
git clone https://github.com/KytraScript/ComfyUI_MatAnyone_Kytra.git
- Install pip requirements
cd ComfyUI/custom_nodes/ComfyUI_MatAnyone_Kytra
pip install -r requirements.txt
- The model weights are automatically downloaded during your first run from: https://huggingface.co/Mothersuperior/ComfyUI_MatAnyone_Kytra and they will be placed in the ComfyUI_MatAnyone_Kytra/model folder
Outdoor Lighting dance example:
MatAnyone_Kytra_Dance_Example.mp4
Realism with low light:
MatAnyone_KytraScript_Example.mp4
Anime using an AI created animation:
Recording.2025-03-09.144015.mp4
Blue car against a blue sky:
MatAnyone_Kytra_Car_Example.mp4
- video_frames: Video image batch (example workflow uses VHS Video Loader)
- mask: First frame mask (subject rgb(255,255,255))
- warmup_frames: Number of warmup iterations (default: 10)
- erode_kernel: Erosion kernel size (default: 10)
- dilate_kernel: Dilation kernel size (default: 10)
- bg_red & bg_green & bg_blue: Set values for background color for composite video output
- Provided in the repository 'workflow' directory
- Original MatAnyone implementation: https://github.com/pq-yang/MatAnyone
- Paper: MatAnyone: Prompting Any Object for Open-world Matting
This project is licensed under the same license as the original MatAnyone project.