ViTeX-Bench Leaderboard
Video scene text editors on three axes: text correctness, temporal quality, and edit locality. There is no aggregate score.
Loading…
- Pareto front
- Other editor
- Reference
- Ideal
Drag to rotate. Select a star for details.
Leaderboard
Mean scores on the frozen 157-video evaluation split. By default, methods are grouped by Pareto front; within a front the order is random, because its members are equally good. Sort by any metric to rank on that metric alone.
| Loading… |
Bold marks the best ranked value in a column. † Flicker and Warp come from interpolated frames and are not compared.
How it is scored
Thirteen metrics on three axes. One primary metric per axis; the chart uses those three.
Pareto front. A raw editor is on the front unless another raw editor is at least as good on all three primary metrics and better on one. Setting the front aside and repeating gives front 2, front 3, and so on. It is computed in three dimensions, so a front member can look dominated in a two-axis view.
Not ranked. Post-processed outputs and the Source video are shown for reference only. Methods marked † are left out of temporal comparisons.
Submit a result
Anyone can submit. A maintainer reviews each result before it appears here.
- 1
Run the evaluation code on the frozen split of ViTeX-Dataset. It writes
eval.json. - 2
Open a submission issue with
eval.jsonattached. Note any post-processing or frame interpolation. - 3
After review, your method appears in the chart and the table.