MatAnyone-E cuts 4K video matting cost 13x, runs 28fps on an RTX 3090
bdsqlsz · x · 2026-11-08
A team from NTU S-Lab and collaborators introduced MatAnyone-E, targeting the compute bottleneck in ultra-HD video matting.
- Approach: a low-resolution pass first locates hard regions like hair edges; high-precision computation happens only there, while flat background reuses low-res results, paired with a lightweight decoder.
- Results: 4K matting runs 7.4x faster with 1/13 the VRAM; a single GPU handles native 4K at 28+ fps and 1080p at 73 fps — smooth even on consumer cards like the RTX 3090 — while preserving hair-level and alpha detail.
- Project page is live; code to be open-sourced.
More from Multimodal
- 10-minute Sherlock Holmes anime made with ComfyUI and Claude Code, files included — shiwano · 2026-11-08
- Kling 4.0 impresses at video restyling, beating Seedance 2.5 in demo — DavidmComfort · 2026-11-08
- NVIDIA open-weights Nemotron 3.5 Transcribe: 30B-param speech model in 40 languages — dl_weekly · 2026-11-08
- Iridescent Prismatic Refraction: a reusable prompt template for spectral AI art — LudovicCreator · 2026-11-08
- SenseTime open-sources Code2Video: code-driven, editable AI video generation — AIwithGhotai · 2026-11-08
- Dev ships interruptible history podcast generator, 80% coded by Claude Code — Goldenchild123 · 2026-11-08