SAGE: Salient Factor Discovery and Generation with Visual Foundation Representations
SAGE isolates target-specific visual factors, then uses them to generate new examples of discovered subtypes.
The paper frames this as contrastive analysis between a target set and a background set, such as faces with and without eyeglasses. It learns salient and common factors in the spatial latent space of a frozen representation autoencoder, then conditions a diffusion transformer on a reference image’s salient representation. In reported tests, it improves digit subtype recovery on Digits-ImageNet, surfaces eyewear styles and mislabeled images in FFHQ, and separates three retinal OCT diseases using only normal/disease labels. Source: HF Daily Papers' note
The paper frames this as contrastive analysis between a target set and a background set, such as faces with and without eyeglasses. It learns salient and common factors in the spatial latent space of a frozen representation autoencoder, then conditions a diffusion transformer on a reference image’s salient representation. In reported tests, it improves digit subtype recovery on Digits-ImageNet, surfaces eyewear styles and mislabeled images in FFHQ, and separates three retinal OCT diseases using only normal/disease labels. Source: HF Daily Papers' note
score 4