The Hidden Language of Diffusion Models

2024-01-16 · ICLR 2024 poster · anchor

Findings

• IC-1137Stable Diffusion represents certain concepts primarily through specific named exemplars rather than abstract category features
• IC-1138Stable Diffusion simultaneously encodes multiple meanings of homograph concepts in a single representation
• IC-1139Stable Diffusion encodes social biases in its internal concept representations that are not always visually apparent
• IC-1140Stable Diffusion's internal concept representations encode visual and structural similarities (shape, texture, color) that transcend textual semantics