Morning Edition · Wednesday, October 28, 2026Published at 2:21 AM EDT · New York

MatAnyone-E Runs 4K Video Matting at 28.62 Frames per Second on One H200 by Limiting High-Resolution Work to Uncertain Regions

The method claims a 7.4x speedup and 13.1x lower peak memory than MatAnyone 2, while national labs using Meta's open vision models have cut beamline analysis from a month to 15 minutes.

Researchers including Shuyun Wang, Shangchen Zhou and Chen Change Loy posted MatAnyone-E, an efficient extension of the MatAnyone 2 video-matting model, on October 26. The method first propagates memory at low resolution to find uncertain r…