Shuyun-Wang/MatAnyone-E

Sparse Memory Propagation for Efficient High-Resolution Video Matting

Shuyun Wang^2†  Shangchen Zhou^2†  Peiqing Yang²  Yuekun Dai²  Xin Yu^3*  Chen Change Loy^2*

¹The University of Queensland  ²S-Lab, Nanyang Technological University  ³University of Adelaide
^†Project lead ^*Corresponding authors

MatAnyone-E mattes native 4K video at 28.6 FPS on a single H200 (with torch.compile), 12× faster than MatAnyone 2, using 3.1 GB of GPU memory, while keeping fine details such as hair strands.

🎥 More results on the project page.

---

• [2026.10] Project page and repository online.

• Paper on arXiv
• Inference code and checkpoints
• Hugging Face demo
• ComfyUI custom node

MatAnyone-E keeps the architecture of MatAnyone 2 and replaces the two parts whose cost grows fastest with resolution:

• Sparse Memory Propagation. A low-resolution pass picks the uncertain query locations (query pruning) and the memory most relevant to them (memory pruning); high-resolution matching runs only on those pairs.
• Lightweight decoder. Designed from stage-wise latency profiling and trained by feature distillation from the original decoder.

FPS and peak GPU memory on one NVIDIA A800-SXM4-80GB (fp16); (C.) = with torch.compile.

~ extrapolated (a full memory bank does not fit in 80 GB). On one H200 with torch.compile: 73.07 FPS at 1080p, 28.62 FPS at 4K.

If you find our repo useful for your research, please consider citing our paper:

@article{wang2026matanyonee,
  title   = {{MatAnyone-E}: Sparse Memory Propagation for Efficient High-Resolution Video Matting},
  author  = {Shuyun Wang and Shangchen Zhou and Peiqing Yang and Yuekun Dai and Xin Yu and Chen Change Loy},
  journal = {arXiv preprint arXiv:2610.35724},
  year    = {2026}
}

This project is licensed under NTU S-Lab License 1.0. Redistribution and use should follow this license.

This project is built upon MatAnyone 2 and MatAnyone. Thanks for their awesome works!

If you have any questions, please feel free to reach us at shuyun.wang@uq.edu.au.