Cross-Attention Head Position Patterns Can Align with Human Visual Concepts in Text-to-Image Generative Models

TL;DR

This paper introduces Head Relevance Vectors (HRVs) to interpret cross-attention layers in diffusion models, enabling better alignment with human visual concepts and improving control over image generation tasks.

Contribution

It presents a mechanistic interpretability method using HRVs to understand and manipulate cross-attention heads in diffusion models, advancing fine-grained control and interpretability.

Findings

HRVs align with human visual concepts in diffusion models

HRVs reduce misinterpretations of polysemous words in image generation

Methods improve attribute editing and multi-concept generation

Abstract

Recent text-to-image diffusion models leverage cross-attention layers, which have been effectively utilized to enhance a range of visual generative tasks. However, our understanding of cross-attention layers remains somewhat limited. In this study, we introduce a mechanistic interpretability approach for diffusion models by constructing Head Relevance Vectors (HRVs) that align with human-specified visual concepts. An HRV for a given visual concept has a length equal to the total number of cross-attention heads, with each element representing the importance of the corresponding head for the given visual concept. To validate HRVs as interpretable features, we develop an ordered weakening analysis that demonstrates their effectiveness. Furthermore, we propose concept strengthening and concept adjusting methods and apply them to enhance three visual generative tasks. Our results show that HRVs can reduce misinterpretations of polysemous words in image generation, successfully modify five challenging attributes in image editing, and mitigate catastrophic neglect in multi-concept generation. Overall, our work provides an advancement in understanding cross-attention layers and introduces new approaches for fine-controlling these layers at the head level.