DeOverlay: A Benchmark and Source-Anchored Flow for Video Overlay Removal
Deshui Miao, Yuhang Ma, Shaoyong Jia, Wenzhao Gao, Yue Zhang, Shaohui Jiao, Xin Li
ByteDance · Southern University of Science and Technology
Published October 19, 2026 · on the feed October 20, 2026 · ▲ 19 · arXiv 2610.26081
The abstract
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. We propose DeOverlay, a source-anchored flow model that reconstructs the clean video from the source latent with fixed source guidance. Specifically, we train a shared velocity network along paired source-to-clean paths using region-weighted flow matching, enabling successive cor- rections from intermediate states. To select local source cues, we develop visibility-normalized temporal aggregation that excludes masked positions and averages the remaining observations. Global source context and edit-region embeddings complement these cues through additive condi- tioning, retaining a single video-token sequence. 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. Additional qualitative examples demonstrate the removal of graphic logos and stickers. We will make the complete benchmark, evaluation code, and test outputs availabl. Project page:https://yahooo-m.github.io/DVTE-Bench/
The Brief
2:29A short podcast summary of the paper. Written and voiced by AI.
Transcript
DeOverlay. A Benchmark and Source-Anchored Flow for Video Overlay Removal. By Deshui Miao and six others, from ByteDance and the Southern University of Science and Technology. Published on the nineteenth of October. No code repository is linked to it.
Video overlay removal means recovering the parts of a scene hidden behind subtitles, titles and logos, while leaving everything around them intact. Existing methods take one of two routes: they propagate content across time, or they generate new content conditioned on the source video. Both can fall short when the useful information in the source is itself obscured. They may leave remnants of the overlay behind, and they may distort the structure of the scene.
DeOverlay is a flow model anchored on the source. It reconstructs the clean video starting from the latent of the source video, with source guidance that stays fixed throughout. Rather than conjuring a clean video from nothing, it travels from the overlaid version to the clean one.
A single shared velocity network is trained along paired paths that lead from source to clean video, using flow matching weighted by region. Because it learns the whole path, it can make successive corrections from intermediate states, not just one jump.
Local cues from the source are chosen by visibility-normalised temporal aggregation. Masked positions are excluded, and the remaining observations across time are averaged, so only content that was actually visible contributes.
Alongside those local cues, global context from the source and embeddings of the region being edited are added in through additive conditioning. That keeps the model working on a single sequence of video tokens.
To evaluate scene recovery across many kinds of overlay, the authors also build DVTE-Bench. It holds one thousand six hundred and thirty-one paired videos across eight categories of text overlay, with clean targets screened by OCR, pixel masks aligned frame by frame, and a common evaluation protocol.
On the complete release, DeOverlay outperforms four baselines on all five reconstruction and perceptual metrics. Against the strongest of those baselines, it reduces Mask LPIPS by fifty-eight point three percent and Mask VFID by seventy point three percent. Further qualitative examples show it removing graphic logos and stickers as well. The authors say they will release the complete benchmark, the evaluation code, and the test outputs.
Two contributions arrive together. One is a model that treats removal as a journey from the overlaid video to the clean one, steered at every step by what the source still shows. The other is a shared test bed with paired ground truth, so that claims about overlay removal can be measured on common terms.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free