ComfyUI-MatAnyone2

ComfyUI-MatAnyone2

A standalone ComfyUI custom node project that wraps MatAnyone2 for video matting.

This repository is independent:

• It includes the matanyone2 inference code (no dependency on your local MatAnyone2 folder).
• It provides ComfyUI nodes for video matting (frame sequence + first-frame mask).
• It supports automatic first-time checkpoint download (matanyone2.pth).

Demo Videos

| Original | Result | | --- | --- | | <video src="https://github.com/user-attachments/assets/14df1772-eb9f-46b5-b4b2-e069332b5d5c" controls muted></video> | <video src="https://github.com/user-attachments/assets/c2517188-2d51-466c-8a95-3ea293312a54" controls muted></video> |

Nodes

1. MatAnyone2 Model Loader

• Loads MatAnyone2 checkpoint weights.
• Can auto-download the checkpoint if it does not exist locally.

2. MatAnyone2 Video Matting

• Inputs: IMAGE frame batch, MASK (first-frame target mask)
• Outputs:
  • foreground: extracted foreground (black background)
  • composite: green-screen composite preview (same style as official inference visualization)
  • alpha: per-frame alpha matte (MASK)

Installation

Clone this repository into ComfyUI/custom_nodes:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/starsFriday/ComfyUI-MatAnyone2.git

Install dependencies:

cd ComfyUI-MatAnyone2
pip install -r requirements.txt

Restart ComfyUI.

Checkpoint

Default checkpoint filename: matanyone2.pth

Download URL:

• https://github.com/pq-yang/MatAnyone2/releases/download/v1.0.0/matanyone2.pth

Recommended local path:

• ComfyUI/models/matanyone2/matanyone2.pth

You can also enable auto_download=true in MatAnyone2 Model Loader to download automatically on first run.

Workflow Usage

1. Prepare a video frame sequence as an IMAGE batch.

• You can use a video loader node to produce IMAGE frames.

2. Prepare the first-frame target mask as MASK.

• White area = foreground person, black area = background.
• If multiple mask frames are provided, only the first one is used.

3. Connect nodes:

• MatAnyone2 Model Loader -> MatAnyone2 Video Matting.matanyone2_model
• video frames -> MatAnyone2 Video Matting.images
• first-frame mask -> MatAnyone2 Video Matting.first_mask

4. Key parameters:

• warmup: first-frame warmup iterations (default: 10)
• erode_kernel / dilate_kernel: first-frame mask morphology
• mask_threshold: mask binarization threshold
• max_internal_size: internal shortest-side limit (-1 = disabled)

Project Structure

ComfyUI-MatAnyone2/
├── __init__.py
├── nodes.py
├── requirements.txt
├── README.md
├── LICENSE_MatAnyone2.txt
├── pretrained_models/
│   └── .gitkeep
└── matanyone2/
    ├── config/
    ├── inference/
    ├── model/
    └── utils/

Notes

• The current node follows the official MatAnyone2 single-target pipeline (objects=[1]).
• This is a human video matting model; result quality strongly depends on input quality and the first-frame mask.
• First model load may be slow; subsequent runs use in-process model cache.

License and Credits

• This project contains and adapts code from MatAnyone2.
• See the original MatAnyone2 license in LICENSE_MatAnyone2.txt.
• Please follow the original license terms for use and redistribution.