DeOverlay: A Benchmark and Source-Anchored Flow for Video Overlay Removal

One-line takeaway

ByteDance's DeOverlay offers a benchmark and a source-anchored flow model for removing subtitles, titles and logos from video, using region-weighted flow matching and visibility-normalized temporal aggregation to cut residue and distortion.

字节的 DeOverlay 提出视频去字幕、去标题和台标的基准与方法：以源视频隐变量为锚做流匹配，按区域加权训练，并按可见度聚合时序线索，减少残留和结构变形。