MatAnyone-E: Sparse Memory Propagation for Efficient High-Resolution Video Matting
The authors report an estimated 7.4× inference speedup at 4K, with peak GPU memory reduced to 7.6% of MatAnyone 2’s usage and comparable matting quality.
MatAnyone-E uses low-resolution processing to locate uncertain regions and select relevant information from earlier frames. It reserves high-resolution matching for those regions and reuses low-resolution results elsewhere. A lightweight decoder further reduces the cost of producing detailed mattes. With compilation enabled on a single NVIDIA H200 GPU, the system reaches 73.07 FPS at 1080p and 28.62 FPS at 4K.
HF Daily Papers' note
MatAnyone-E uses low-resolution processing to locate uncertain regions and select relevant information from earlier frames. It reserves high-resolution matching for those regions and reuses low-resolution results elsewhere. A lightweight decoder further reduces the cost of producing detailed mattes. With compilation enabled on a single NVIDIA H200 GPU, the system reaches 73.07 FPS at 1080p and 28.62 FPS at 4K.
HF Daily Papers' note
score 4