SAGE discovers latent variables that, when used to generate images, produce distorted or garbled outputs regardless of the text prompt. These latent variables are non-outlier samples of N(0,I), confirmed by Shapiro-Wilk tests (p-values 0.429, 0.207, 0.534) and mean/std checks. The authors show these are not isolated points but connected regions: expanding the boundary around each found latent variable reveals that any sample within the region produces similar distorted images. The distorted images can be generated with non-zero probability under standard Gaussian sampling, indicating the latent space is not well-structured in these regions.
Evidence
correlational
Key metric
Shapiro-Wilk p-values: 0.429, 0.207, 0.534; mean/std: -0.003/1.005, -0.027/1.045, 0.008/1.041; connected regions confirmed by boundary expansion
Caveat
The distorted images are classified with high confidence by ViT and ViT on ImageNet-C (e.g., pinwheel 0.96, dalmatian 0.99, warplane 0.99), suggesting the distortion aligns with classifier representation biases rather than being purely random noise.