Identifying Models Behind Text-to-Image Leaderboards

TL;DR

This paper demonstrates that anonymized text-to-image model outputs on leaderboards can be deanonymized using image embedding clustering, exposing security flaws and suggesting the need for improved anonymization methods.

Contribution

It introduces a centroid-based method for deanonymizing T2I models and analyzes prompt-level distinguishability, revealing vulnerabilities in current leaderboard practices.

Findings

High accuracy in model deanonymization using embedding clustering

Certain prompts can nearly perfectly distinguish models

Fundamental security flaws in T2I leaderboard anonymization

Abstract

Text-to-image (T2I) models are increasingly popular, producing a large share of AI-generated images online. To compare model quality, voting-based leaderboards have become the standard, relying on anonymized model outputs for fairness. In this work, we show that such anonymity can be easily broken. We find that generations from each T2I model form distinctive clusters in the image embedding space, enabling accurate deanonymization without prompt control or training data. Using 22 models and 280 prompts (150K images), our centroid-based method achieves high accuracy and reveals systematic model-specific signatures. We further introduce a prompt-level distinguishability metric and conduct large-scale analyses showing how certain prompts can lead to near-perfect distinguishability. Our findings expose fundamental security flaws in T2I leaderboards and motivate stronger anonymization defenses.