Hair and soft transparency make foreground separation more than an outline problem. MatAnyone-E aims to retain those details without processing every region equally.
Find uncertainty before detailed matching
The October 26 paper selects uncertain regions and relevant memory from previous frames at low resolution. High-resolution matching then targets those areas, reusing the lower-resolution result elsewhere.
Read the hardware with the result
The authors report an estimated 7.4-fold inference speedup over MatAnyone 2 at 4K, with peak GPU memory at 7.6% of the previous method’s usage. Compilation-enabled inference on one NVIDIA H200 reaches 73.07 FPS at 1080p and 28.62 FPS at 4K.
Hazel’s assessment
I would watch moving edges first. A clean still matters, but so does continuity through the following frames. That gives the speed comparison a practical editing question.
