DeOverlay: A Benchmark and Source-Anchored Flow for Video Overlay Removal
We propose DeOverlay, a source-anchored flow model that reconstructs the clean video from the source latent with fixed source guidance.
ProofPaper ↗
Key points
- Video overlay removal aims to recover scenes obscured by subtitles, titles, and logos while preserving surrounding content.
- To select local source cues, we develop visibility-normalized temporal aggregation that excludes masked positions and averages the remaining observations.
- To evaluate scene recovery across diverse overlay conditions, we introduce DVTE-Bench, comprising 1,631 paired videos across eight text-overlay categories, with OCR-screened clean targets, frame-aligned pixel masks, and a common evaluation protocol.
- Experiments on the complete release show that DeOverlay outperforms four baselines on all five reconstruction and perceptual metrics, reducing Mask LPIPS and Mask VFID by 58.3% and 70.3%, respectively, relative to the strongest evaluated baseline.
Sources (2)
- [1]DeOverlay: A Benchmark and Source-Anchored Flow for Video Overlay RemovalHugging Face Daily Papers · Oct 19, 12:00 AM
We propose DeOverlay, a source-anchored flow model that reconstructs the clean video from the source latent with fixed source guidance.
Video overlay removal aims to recover scenes obscured by subtitles, titles, and logos while preserving surrounding content.
- [2]DeOverlay: A Benchmark and Source-Anchored Flow for Video Overlay RemovalarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 09:39 AM · same content
Extractive summary: sentences quoted from the sources.
Before this
- Oct 18, 2026Test-Time Scaling of Text-to-Image Flow Matching Models with Noise-Rejuvenated Sequential Monte Carlo
- Oct 18, 2026VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse
- Oct 17, 2026CorrFlow: Self-Correcting Flow Matching
- Oct 16, 2026CLEAR: Cooperative Latent Flow for Visual Counterfactual Explanations with Accurate Region Preservation
- Oct 16, 2026TriFold: Folding irregular 3D Gaussians into structured latents for 3D Scene Generation
- Oct 16, 2026Welcome Beam: a 501B open model you can own, customize, and scale