IC-1013Latent samples in Stable Diffusion v2.1, v1.5, and GLIDE can produce images of associated backgrounds rather than the key object, with failure rates of 2.7%, 9.2%, and 50.5% under random sampling respectively

Qihao Liu, Adam Kortylewski, Yutong Bai, Song Bai, Alan Yuille

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.

Evidence
correlational
Key metric
Region size: SD v2.1 0.024, SD v1.5 0.037, GLIDE 0.060; Failure rate (random sampling): SD v2.1 2.7%, SD v1.5 9.2%, GLIDE 50.5%; SSR(a): SD v2.1 32.0%, SD v1.5 41.5%, GLIDE 96.5%
Caveat
The five brittle categories (umbrella, basketball, car, unicycle, alp) were selected by running SAGE on all 1k ImageNet categories and ranking by convergence speed, so they are the worst cases rather than a random sample.
Model
Stable Diffusion v2.1, v1.5, GLIDE
Concepts
Failure mode
Datasets
ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval]
Methods
Shapiro-Wilk test [validation]
Related findings
IC-1012, IC-1014
Extraction
automatic-extraction