Identifying the Models Behind Text-to-Image Leaderboards

Researchers propose a centroid-based deanonymization method that exploits the distinctive clusters formed in image embedding space by images generated by each text-to-image model, identifying the models behind voting-based leaderboards without prompt control or training data. Experiments cover 22 models, 280 prompts, and a total of 150K images; the method achieves high accuracy and reveals systematic model-specific generation characteristics. The paper also proposes a prompt-level distinguishability metric, and large-scale analysis shows that some prompts can make models nearly perfectly distinguishable. The authors point out that these findings expose a fundamental security flaw in text-to-image leaderboards and motivate stronger anonymization defenses.