DVTE-BENCH / ALIGNED PREVIEW01 / 16
Inspect a paired scene
Select a case to compare its frames.
Source could not be loaded.
Comparison could not be loaded.
0.0 s0.0 s
DIVERSE VIDEO TEXT ERASURE BENCHMARK
Erase the text.
Keep the story.
DeOverlay + DVTE-Bench
A closer look at video text erasure. Eight kinds of text, paired clean references, and frame-aligned masks reveal what happens beyond ordinary subtitles.
A RECORDED MODEL COMPARISON
OURS · RECORDED OUTPUTSOURCE
MULTI-REGION · GT BOX · 6 STEPSDrag to reveal
One scene. Every text region.
1,631 clips
Paired source + reference
8 types
Layout, motion, typography
136,090
Frame-aligned video frames
91.1 min
Portrait + landscape video
01 / MORE THAN SUBTITLES
Text takes many forms.
Test all eight.
From moving captions to vertical writing, each slice brings a different challenge. Pick a card to inspect its paired benchmark examples.
Card previews show source clips with text overlays. Hover or focus to play; select to explore.
02 / LOOK CLOSER
A scene. A text layer.
Every frame matters.
Compare the source with recorded model outputs, the paired clean reference, or the text mask. One timeline keeps the views aligned.
Loading the benchmark…
Recorded outputs are qualitative previews. The clean reference is paired benchmark data; it is not a model prediction. Metrics below use the full 1,631-video evaluation.
03 / THE BENCHMARK
Look at the pixels.
Then the numbers.
Five methods, all 1,631 clips. Explore quality and temporal metrics inside the dataset’s pixel-accurate text regions.
04 / WHAT IS MEASURED
Aligned data.
Explicit evaluation.
Each source has a clean reference and a frame-level mask. Category breakdowns keep the different text challenges visible.
Pair the frames.
Source, clean reference, and text masks share frame alignment. Compare the same scene at the same time.
Inspect the text region.
Quality metrics use dataset ground-truth pixel masks. Temporal metrics inspect the affected regions across frames.
Report every slice.
All eight categories contribute to the full aggregate. Per-type results reveal where performance changes.
Clean-target audit. All target frames were scanned under one OCR configuration. 569 candidates with confirmed residual text were rejected before the final release.
Dataset composition. 1,563 clips use synthetically overlaid text; 68 form the seen ASR regression track with known training overlap. The full 1,631-video aggregate and category results retain both tracks.
Inference settings. Ours uses a 40K checkpoint, 6 inference steps, GT bounding-box guidance, and seamless paste-back. ProPainter, MiniMax-Remover, and DiffuEraser use GT box masks with source paste-back; CLEAR is mask-free. Metrics use pixel-accurate GT masks. The qualitative outputs include each method’s saved postprocessing.
A CLOSER LOOK AT THE DATA
66.5%PORTRAIT
1,085 portrait + 546 landscape clips
11language labels
across the benchmark
across the benchmark
05 / EXPLORE THE DATA
Find your next
stress test.
Browse all 1,631 records, inspect the metadata, or take the manifests into your own evaluation workflow.
Search the complete benchmark 1,631 records · type, language, format
| Sample ID | Type | Language | Format | Duration | Regions |
|---|
DVTE-BENCH / PAIRED DATASource. Reference. Mask.
Source. Reference. Mask.
Ready to inspect.
Download the complete metadata or explore the source repository.