Localizing and Editing Knowledge in Text-to-Image Generative Models
This addresses the problem of efficiently editing concepts in text-to-image models for researchers and practitioners, offering a significant speed improvement but is incremental as it builds on existing causal analysis and editing techniques.
The paper investigates how knowledge about visual attributes is stored in text-to-image diffusion models, finding that it is distributed across UNet components and localized in a single causal state in the text-encoder. Based on this, they introduce Diff-QuickFix, a fast editing method that achieves a 1000x speedup with comparable performance to fine-tuning methods.
Text-to-Image Diffusion Models such as Stable-Diffusion and Imagen have achieved unprecedented quality of photorealism with state-of-the-art FID scores on MS-COCO and other generation benchmarks. Given a caption, image generation requires fine-grained knowledge about attributes such as object structure, style, and viewpoint amongst others. Where does this information reside in text-to-image generative models? In our paper, we tackle this question and understand how knowledge corresponding to distinct visual attributes is stored in large-scale text-to-image diffusion models. We adapt Causal Mediation Analysis for text-to-image models and trace knowledge about distinct visual attributes to various (causal) components in the (i) UNet and (ii) text-encoder of the diffusion model. In particular, we show that unlike generative large-language models, knowledge about different attributes is not localized in isolated components, but is instead distributed amongst a set of components in the conditional UNet. These sets of components are often distinct for different visual attributes. Remarkably, we find that the CLIP text-encoder in public text-to-image models such as Stable-Diffusion contains only one causal state across different visual attributes, and this is the first self-attention layer corresponding to the last subject token of the attribute in the caption. This is in stark contrast to the causal states in other language models which are often the mid-MLP layers. Based on this observation of only one causal state in the text-encoder, we introduce a fast, data-free model editing method Diff-QuickFix which can effectively edit concepts in text-to-image models. DiffQuickFix can edit (ablate) concepts in under a second with a closed-form update, providing a significant 1000x speedup and comparable editing performance to existing fine-tuning based editing methods.