SourceLocalizing and Editing Knowledge In Text-to-Image Generative Models
Using causal tracing (corrupting subject token embeddings and restoring individual UNET layers), the paper identifies where knowledge about objects, style, action, and color resides in Stable Diffusion's 70-layer UNET. Unlike LLMs where knowledge localizes to a few mid-MLP layers, causal states in the UNET are spread across diverse components. The distribution is attribute-specific: self-attention-0 is causal for style but not for objects, action, or color; mid-block cross-attention is causal for action but not other attributes. Notably, cross-attention layers are not the primary causal states; resnet blocks and self-attention blocks carry significant knowledge. The density of causal states is higher in the up-block than in the down-block or mid-block.