Text Refinement Bench
A benchmark for evaluating scene-text refinement in generated images, measuring correction accuracy, background preservation, and efficiency using character-level error annotations.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info text-refinement-bench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
2 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| PixText | 83.1 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen-Image-3.0-Pro | 28.2 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
Constructed from model-generated images with character-level error annotations and correction targets. It evaluates localized text correction capabilities rather than general text generation.
Scoring focuses on glyph pass rate and exact target-character OCR accuracy. PixText achieves 83.06% glyph pass and 94.66% OCR accuracy, significantly outperforming baselines.