StreamErase hits 26 FPS real-time video object erasing on a single GPU, 200x faster than prior methods
Zhuobai Dong · hf · 2026-10-15
Researchers from Wuhan University, Alibaba, HKUST and SJTU present StreamErase, cutting video object removal latency from minutes (e.g. VACE at 204s) to 0.87s at up to 26.06 FPS on one GPU.
Key ideas:
- A bidirectional text-to-video teacher with mask input learns to erase objects and synthesize coherent backgrounds.
- The erasing capability is progressively distilled into a unidirectional streaming student model.
- Sliding-window causal attention with attention sinks, plus self-rollout and KV cache rolling during training, simulate inference-time extrapolation with a fixed context window, preventing error accumulation and enabling constant-speed generation of arbitrarily long videos.
The team also releases LongErase, a benchmark for streaming long-video object erasure. The model matches or beats existing erasing models and uniquely supports interactive, real-time masking with infinite-length streaming.
More from Multimodal
- Spooky season: painterly AI Halloween illustrations with film-poster vibes — egeberkina · 2026-10-15
- Classic artworks reimagined as animated pixel art with Gemini 4 — fofrAI · 2026-10-15
- Glance Forcing cuts video distillation trajectory cost from thousands of GPU-hours to one — Zhuobai Dong · 2026-10-15
- Skywork open-sources Matrix-Game 4.0, a playable world model running in real time on one RTX 5090 — mishig25 · 2026-10-15
- Alibaba TaoLive open-sources 3-step LoRA for real-time streaming audio-video generation — linoy_tsaban · 2026-10-15
- Entire video generated by Opus 5.5 on Max reasoning, no skills used — Hesamation · 2026-10-15