VDE Bench (Local) - BLEU-4: leaderboard
Metric: BLEU-4 of the OCR text against the gold text over matched blocks (0-1, shown times 100), within the single annotated edit region (the output cropped to the gold edit box); VDE Bench's 942 human-verified instruction-based edits of dense Chinese and English document pages (papers, slides, posters, exams, newspapers; text replacement, insertion and deletion and table edits; gold edited pages verified by 15 annotators); the edited page is read with PaddleOCR-VL and compared with the gold page, default generation settings; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | FireRed-Image-Edit-1.0 | 33.6 | |
| 2 | Qwen-Image-Edit (VDE Bench checkpoint unspecified) | 32.8 | |
| 3 | LongCat-Image-Edit | 32 | |
| 4 | Step1X-Edit (VDE Bench checkpoint unspecified) | 17.4 | |
| 5 | ICEdit | 8.5 | |
| 6 | InstructPix2Pix | 7.8 |
Interactive version: theaggregate.ai/benchmark?slug=vde-bench-local-bleu-4 · How It Works · Data refreshed daily, snapshot 2026-10-11.