ByteDance's DeOverlay: Source-Anchored Flow Matching for Video Overlay Removal with New Benchmark
ByteDance · hf · 2026-10-20
ByteDance and SUSTech release DeOverlay, a model that removes video overlays like subtitles, titles, and logos while preserving scene structure.
- A source-anchored flow model reconstructs clean video from the source latent with fixed source guidance
- A shared velocity network is trained along paired source-to-clean paths using region-weighted flow matching, enabling successive corrections from intermediate states
- Visibility-normalized temporal aggregation selects local source cues
- Introduces DVTE-Bench for standardized evaluation of overlay removal
More from Multimodal
- MOSS Image with 8-step Turbo LoRA tested: 10x faster with minimal quality loss — Aggressive_Collar135 · 2026-10-20
- User turns their AI muse Maki into a punky ska song with Suno V6 — thederbiedone · 2026-10-20
- Grok Imagine parody short 'Wendy' shows off AI-made TV skits — Old_Put1306 · 2026-10-20
- AI-generated "Back to the Future" sequel clip goes viral for Halloween — mitchdeg · 2026-10-20
- Midjourney v8.2 style ref 1631327937 yields a dreamy pastel look — azed_ai · 2026-10-20
- Google Flow hosts hybrid production workshop pairing live-action with AI filmmaking workflows — lmoroney · 2026-10-20