ComfyUI-MatAnyone

★ 86

视频抠像去背景高质量连贯性增强

ComfyUI-MatAnyone 节点用于在 ComfyUI 中运行 MatAnyone 系列的视频抠像，去除背景并输出透明或纯色前景，支持 MatAnyone/MatAnyone2 提升视频连贯性与质量。

💡 将视频人物抠出为透明通道或纯色背景以便后期合成。

🍴 5 Forks💻 Python🔄 2026-03-15

🔗 GitHub 原文

📦

网盘下载

复制链接后前往夸克网盘下载

https://pan.quark.cn/s/e58c8376a81b

📦 requirements.txt

torch
omegaconf

📄 README

ComfyUI-MatAnyone

MatAnyone in ComfyUI (Remove background)

Stable Video Matting with Consistent Memory Propagation:

Scaling Video Matting via a Learned Quality Evaluator:

Download matanyone.pth from

Download matanyone2.pth from

checkpoint/
    matanyone.pth
    matanyone2.pth

https://github.com/user-attachments/assets/f1805028-92ed-40b7-8bf5-85f1a3ddbe25

Workflow

The extension now includes the MatAnyone2Video node, which runs the improved MatAnyoneV2 model for higher-quality and more robust video matting.

Workflow: workflow/workflow_mat_anyone.json

(Not a workflow-embedded image)

Inputs:

src_video

forground_mask (IMAGE) or foreground_MASK (MASK): The input mask. IMAGE option will automatically convert a black/white image to a mask. At least one option must be given.

solid_color (optional): The solid color to create a screen. Defaults to Green Screen.

mask_frame: The input mask’s index (defaults to 0). Support first (0), last and middle frame.

n_warmup: Number of iterations to warm up the model. Defaults to 10.

Memory Management Inputs:

max_internal_size (optional): Resizes the internal processing resolution to save memory (e.g., 360, 480). Default is -1 (full resolution).

max_mem_frames (optional): The number of key frames kept in high-resolution working memory. Default is 5.

use_long_term (optional): Limits active memory scaling by compressing older frames into long-term prototype memories, preventing OOM over time. Default is False. (Useful for long videos)

Your input mask won’t actually be in the final matte. Instead, the warmup process generate a new input mask, which is then propagated throughout the video.

Additional Inputs for V2:

r_erode (optional): The radius for morphological erosion applied to the foreground_mask before processing (defaults to 0). Useful for refining rough masks.

r_dilate (optional): The radius for morphological dilation applied to the foreground_mask before processing (defaults to 0).

Credit

@InProceedings{yang2025matanyone,
    title     = {{MatAnyone}: Stable Video Matting with Consistent Memory Propagation},
    author    = {Yang, Peiqing and Zhou, Shangchen and Zhao, Jixin and Tao, Qingyi and Loy, Chen Change},
    booktitle = {arXiv preprint arXiv:2501.14677},
    year      = {2025}
}

@InProceedings{yang2026matanyone2,
   title     = {{MatAnyone 2}: Scaling Video Matting via a Learned Quality Evaluator},
   author    = {Yang, Peiqing and Zhou, Shangchen and Hao, Kai and Tao, Qingyi},
   booktitle = {CVPR},
   year      = {2026}
}