MatAnyone-E: 7.4x faster 4K video matting with 7.6% of the GPU memory
ccloy · x · 2026-10-27
Researchers from S-Lab (NTU), University of Queensland and Adelaide present MatAnyone-E, an efficient extension of MatAnyone 2 for high-resolution video matting.
- 7.4× faster inference at 4K, peak GPU memory cut to 7.6% (3.1 GB) of MatAnyone 2, with comparable matting quality
- With torch.compile, it hits 73 FPS at 1080p and 29 FPS at 4K on a single NVIDIA H200 — 12× faster than MatAnyone 2 compiled
- Method: low-resolution propagation first identifies uncertain regions and relevant memory candidates; high-resolution matching runs only there while reusing low-res results elsewhere, plus a lightweight decoder designed around measured runtime bottlenecks
- Preserves fine details like hair strands and soft transparency with frame consistency
Paper on arXiv; code coming soon.
More from Multimodal
- SATE-Bench: 316 scenes show MLLMs stop exploring too early with high confidence — Tianchen Guo · 2026-10-27
- Two-year-old patch still fixes ComfyUI's slow checkpoint loading: 225MB/s to 2.5GB/s — Solongtomegrandma · 2026-10-27
- A lava-slide first-person video made with Dreamina for just 29 cents — Certain_Menu_8412 · 2026-10-27
- Full prompt revealed for AI-generated real estate tour videos with timestamped shots and lip sync — FinanceYF5 · 2026-10-27
- First working AI real estate video workflow: Kling 4 one-shot with consistent agent and house — FinanceYF5 · 2026-10-27
- 40-person blind test: people spot AI video at 52%, barely better than a coin flip — Resident-Anxiety6427 · 2026-10-27