Diverse Video Text Erasure Benchmark

DVTE-Bench

1,631 paired clips, 136,090 frames, and 8 text types spanning the layouts that challenge conventional video text erasers.

View results Samples Manifest

Paired supervision Frame-level masks OCR-audited targets Portrait + landscape

Live benchmark sample Source with artistic title

1,631 Paired clips Source, target, and mask

8 Text types Layout, motion, and style stressors

136,090 Video frames Frame-aligned evaluation data

91.1 Minutes All clips include audio

Overview

Evaluation beyond bottom-aligned subtitles

DVTE-Bench isolates difficult cases hidden by aggregate scores: moving overlays, vertical writing, distributed regions, title cards, persistent marks, and non-bottom layouts.

Every source clip has a clean target and a temporally aligned mask. Target frames were scanned under one OCR audit configuration; confirmed residual text caused rejection. All eight text types, including ASR subtitles, contribute to the same 1,631-video aggregate.

Taxonomy

Eight complementary stress tests

Each slice targets a distinct spatial, temporal, or typographic failure mode.

Benchmark Results

Mask-region leaderboard

Five methods evaluated on all 1,631 clips inside pixel-accurate dataset ground-truth masks.

Breakdown by text type

Standard Subtitle · 232 videos · pixel mask region

Text type

† Finite Mask PSNR mean; zero-MSE videos are excluded only from that column. All quality metrics use pixel-accurate masks. Mask VFID is a global full-benchmark metric and is omitted from per-type breakdowns.

Interactive Explorer

Inspect aligned evidence

Select a case and compare source, clean target, and mask at the same timestamp.

16 web previews

Synchronized preview 01 / 16

Source

Clean target

Alpha mask

0:00 0:00

Standard Subtitle · Landscape

Loading benchmark case

Composition

Balanced across diverse challenges

All 1,631 videos contribute to the aggregate, with results also broken down by text type.

1,085 Portrait clips 66.5% of the benchmark

546 Landscape clips 33.5% of the benchmark

11 Language labels Chinese, English, Korean, Japanese, Thai, and long-tail ASR

Protocol

A reproducible evaluation contract

Evaluate every clip consistently against its paired clean reference and pixel-accurate mask.

Load paired clips

Read source and target at native frame alignment. Use the provided mask for region metrics.

Report mask region

Measure Mask PSNR, SSIM, LPIPS, DISTS, VFID, TWE, TC, and flow statistics.

Evaluate all clips

Compute the primary leaderboard over the complete set of 1,631 benchmark videos.

Break down by type

Report all eight text categories separately to expose category-specific strengths and failures.

Catalog

Search all 1,631 records

Sample	Type	Language	Format	Duration	Regions

Page 1

Data Package

Structured for direct evaluation

Each type repeats the same source, target, mask, and per-sample metadata layout.

CSV manifest Web manifest

multitype_synth_v4_vlm_asr_ocr_clean_1631/

metadata.csv
summary.json
<benchmark_type>/
  metadata.csv
  source/        # text-overlaid MP4
  target/        # clean reference MP4
  mask_alpha/    # visual mask MP4
  mask_npz/      # frame-aligned masks
  sample_meta/   # overlay provenance

Metadata SHA-256 58ef4419...ef965