nanobench

Navigation

How it works All evals My evals

Legal

Terms of service Privacy policy

Impossible Prompts (version 2)

Checklist
5 prompts
·
13 models
Raters check:
  • The image accurately depicts: "{prompt}"
  • The image is high quality and free of artefacts
Completed
Ratings CSV One row per individual rating Results JSON Eval settings, scores, and every rating Images (zip) Kept for at least 5 years after creation
Clone eval

Add more responses

More responses tighten the confidence intervals on each model's accuracy.
Total
additional responses

Accuracy

% of statements checked true

Share of statements checked true. Hover for exact counts.

Cost vs. accuracy

True rate by statement

Statement GPT-5.4 Image 2 GPT Image 2 GPT Image 1.5 (High Fidelity) Nano Banana Pro (Gemini 3 Pro) Flux.2 Max Nano Banana 2 (Gemini 3.1 Flash) GPT Image 1 Mini Krea 2 Medium Flux.2 Pro Seedream 4.0 Flux.1 Schnell Flux.2 Dev Turbo Flux.2 Klein 9B
The image accurately depicts: "{prompt}"
Global
 87% (15) 80% (15) 60% (15) 47% (15) 33% (15) 47% (15) 33% (15) 53% (15) 33% (15) 20% (15) 7% (15) 7% (15) 0% (15)
The image is high quality and free of artefacts
Global
 93% (15) 93% (15) 93% (15) 93% (15) 93% (15) 80% (15) 93% (15) 73% (15) 80% (15) 80% (15) 87% (15) 87% (15) 87% (15)