SciFig-Bench - Reference-Based Fidelity: leaderboard
SciFig-Bench (Huang et al., 2026) asks a system to draw the methodology figure of an AI or ML paper from its method text and a figure description, over 435 human-verified papers from 15 venues. Reference-based fidelity (0-10): the mean of DINOv2 visual similarity, SigLIP2 figure-text relevance and OCR text F1 against the figure the authors drew; higher is better.
Metric: Reference-based fidelity (0-10) (mean of DINOv2 visual similarity, SigLIP2 figure-text relevance and OCR-F1 against the author-drawn figure). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Nano Banana Pro | 3.4 | |
| 2 | SciFig GPT-5.4-Image (checkpoint unspecified) | 2.9 | |
| 3 | Qwen-Image | 2.1 |
Interactive version: theaggregate.ai/benchmark?slug=scifig-bench-reference-based-fidelity · How It Works · Data refreshed daily, snapshot 2026-11-12.