AtelierEval - Image Constraint Satisfaction (Skilled Prompting): leaderboard
AtelierEval (Luo et al., 2026) tests models as text-to-image prompters over 360 expert-written tasks. Each prompt is rendered four times by each of four image generators (Gemini 3 Pro Image, GPT-Image-1, Flux.1 Pro, SDXL), the best image per generator is kept, and a GPT-5.4 judge scores the share of the constraint checklist of the task visible in it, averaged over generators and task categories, in percent (higher is better). The model receives a structured system prompt asking for chain-of-thought analysis and self-verification.
Metric: Image checklist satisfaction (%; share of each task's expert constraint checklist realized in the generated image, best of four images per backend, mean over four text-to-image backends (Gemini 3 Pro Image, GPT-Image-1, Flux.1 Pro, SDXL) and the three task categories, GPT-5.4 AtelierJudge; structured prompting with chain-of-thought and self-verification). Source: arxiv.org. Saturation forecast: Around 2029. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 65.4 | #172 |
| 2 | Gemini 3 Pro (Preview) | 65.4 | #111 |
| 3 | Claude Sonnet 4.5 | 65.1 | #244 |
| 4 | GPT-4.1 | 63.9 | #492 |
| 5 | Qwen 3 VL 235B A22B | 63.1 | #477 |
| 6 | Gemini 2.0 Flash | 62.9 | #702 |
| 7 | GPT-4.1 Nano | 62.8 | #1992 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ateliereval-image-constraint-satisfaction-skilled-prompting · How It Works · Data refreshed daily, snapshot 2026-11-12.