MatAnyone-E makes 4K video matting 7.4x faster using just 7.6% of GPU memory
ccloy · x · 2026-10-27
MatAnyone-E delivers high-quality video matting 7.4x faster than MatAnyone 2 at 4K, using only 7.6% of peak GPU memory (73 FPS at 1080p, 29 FPS at 4K on one H200). The trick is selective computation: a low-res pass first, refining only uncertain regions (5.9% of query locations in the demo), searching only the most relevant memory slots (15% of candidates), skipping 99.1% of high-res matching pairs. Paper and project page are out; code coming soon.
Related event: MatAnyone-E Speeds Up 4K Video Matting 7.4x With Only 7.6% Memory(3 posts)→
More from Multimodal
- Seedance 2.5 + Suno V6 + GPT team up for a meme MV remix: "More RAM, less scam!" — ivan_bezdomny · 2026-10-28
- Testing MOSS Image in ComfyUI: 9 free workflows for text-to-image and in-image text — pixaromadesign · 2026-10-28
- Genie 4 ships with SynthID watermarking, coming to Google AI Ultra subscribers via Project Genie — GoogleAI · 2026-10-28
- DeepMind unveils Genie 4, its most realistic and interactive world model yet — GoogleDeepMind · 2026-10-28
- MiniMax H3 refmod labels inconsistently switch between Picture and Video references — Azhram · 2026-10-28
- Black Forest Labs lands $300M Series B as CEO pitches open visual models for robotics — bfl_ai · 2026-10-28