TangPoetryBench: leaderboard
TangPoetryBench (Hu et al., 2026) collects human ratings of text-to-image illustrations for 320 classical Chinese Tang poems on ten dimensions, from technical quality and text integrity to core imagery and emotional resonance. Score is the per-image mean of the applicable dimensions on a 0-100 scale; higher is better.
Metric: Human quality score (0-100): per-image mean over the applicable of ten rated dimensions, from visual quality and poem correspondence to imagery, emotion and overall impression, averaged across raters and over 320 poem illustrations. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Nano Banana Pro | 91.1 | |
| 2 | GPT Image 1 | 89.2 | |
| 3 | Seedream 4.5 | 88 | |
| 4 | Midjourney V7 | 72.5 |
Interactive version: theaggregate.ai/benchmark?slug=tangpoetrybench · How It Works · Data refreshed daily, snapshot 2026-11-12.