SciDiagramEdit - Image Generation Semantic Win Rate: leaderboard
SciDiagramEdit paper v1 test split from 364 author-revision figure pairs. Semantic and aesthetic win rates use blind pairs against the author's revised figure, judged by Claude Opus 4.7. The train/validation/test split is 2:1:3. Native fractions are retained; superscript improvements are not the score. Direct raster image generation; no code editor or skill evolution. Table 2 condition: Image Generation. Checklist success is judged against atomic edit requirements. IAA and ISTA are UniPercept scores on a 0–100 scale; pairwise win rates and success are fractions. Metric: Rate (0–1).
Metric: Rate (0–1). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-Image-2 | 0.78 | |
| 2 | GPT-Image-1.5 | 0.35 |
Interactive version: theaggregate.ai/benchmark?slug=scidiagramedit-image-generation-semantic-win-rate · How It Works · Data refreshed daily, snapshot 2026-10-26.