SourceDiscovering Failure Modes of Text-guided Diffusion Models via Adversarial Search
SAGE identifies latent variables that generate natural-looking images depicting associated backgrounds (e.g., a road for 'car', a beach for 'umbrella') instead of the key object. These latent variables are non-outlier samples of N(0,I). The authors define 'model stability' as the probability of generating an irrelevant image under random sampling and find it varies dramatically across models: GLIDE has a 50.5% failure rate on five brittle categories, SD v1.5 has 9.2%, and SD v2.1 has 2.7%. The normalized size of the failure region correlates with the random-sampling failure rate, suggesting the type-2 failure region size is a direct measure of model stability.