DeOverlay reconstructs clean video by anchoring flow to source latents
Original title: DeOverlay: A Benchmark and Source-Anchored Flow for Video Overlay Removal
DVTE-Bench covers 1,631 paired videos; DeOverlay lowers Mask LPIPS 58.3% and Mask VFID 70.3% versus the strongest baseline.
DeOverlay is a source-anchored flow model for removing subtitles, titles, logos, and stickers while keeping surrounding scene structure. A shared velocity network is trained on paired source-to-clean paths with region-weighted flow matching, plus visibility-normalized temporal aggregation. On the authors' complete DVTE-Bench release it beats four baselines on all five reconstruction and perceptual metrics. They say the benchmark, evaluation code, and test outputs will be released.
Unconfirmed: Public release of the full benchmark and code is stated as forthcoming, not confirmed here.
- The benchmark spans eight text-overlay categories with OCR-screened clean targets.
- Visibility-normalized aggregation drops masked positions before averaging remaining observations.
- Global source context and edit-region embeddings are added while keeping one token sequence.
- Qualitative examples also remove graphic logos and stickers, not only text.
From the original article
1 min read at huggingface.co
Video overlay removal aims to recover scenes obscured by subtitles, titles, and logos while preserving surrounding content. Existing methods use temporal propagation or source-conditioned generation, but can leave residual overlays and distort scene structure when useful source information is obscured.
By Deshui Miao, Yuhang Ma, Shaoyong Jia · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.