Study Shows Marginal Matching Fails to Prevent Style Leakage in Factorized Generative Models

A recent study published on arXiv (2608.05243) questions a prevalent belief regarding factorized generative models, commonly utilized in AI art and image creation. The researchers reveal that aligning the marginal distribution of a latent style variable with a Gaussian prior does not ensure independence from class data. They present a precise decomposition indicating that this marginal alignment is merely one of four necessary conditions for effective factorized sampling, highlighting that resolving the mismatch is essential but insufficient. Their empirical analysis shows that both their case-study model and four baseline models achieve nearly zero global Maximum Mean Discrepancy (MMD), yet a linear probe can still retrieve class labels with 74%–100% accuracy, significantly above the 10% chance level. These results raise concerns about the reliability of style transfer and controlled generation in AI art systems, as unintended class leakage may result in biased or unpredictable outputs.