MatAnyone-E

MatAnyone-E is a video matting model that separates foreground from background while preserving fine detail such as hair strands and soft transparency, with temporal consistency across frames. It is an efficient extension of MatAnyone 2, developed by Shuyun Wang, Shangchen Zhou, Peiqing Yang, Yuekun Dai, Xin Yu and Chen Change Loy.\n\nIt keeps the MatAnyone 2 architecture but replaces the two components whose cost grows fastest with resolution. Sparse Memory Propagation runs a low-resolution pass to find uncertain regions and the most relevant memory, then performs high-resolution matching only on those query–memory pairs. A lightweight decoder, trained with feature distillation from the original decoder, reduces the cost of producing high-resolution mattes.\n\nThe project page reports matting quality comparable to MatAnyone 2 (e.g. VideoMatte HR 4.14 vs 4.10) with an estimated 7.4× inference speedup and peak GPU memory reduced to 7.6% at 4K. With torch.compile on a single NVIDIA H200 it reaches 73.07 FPS at 1080p and 28.62 FPS at 4K, using about 3.1 GB of GPU memory. Based on the provided material, code and a ComfyUI node were described as forthcoming; no public code or weights link was found, so availability and licence are unconfirmed.