Unveiling Concept Attribution in Diffusion Models

TL;DR

This paper introduces CAD, a framework for understanding how components in diffusion models contribute to concept generation, revealing both positive and negative influences, and enabling targeted model editing.

Contribution

The paper presents a novel component attribution method for diffusion models that uncovers both positive and negative concept-related components, enhancing interpretability and enabling precise model editing.

Findings

Identifies both positive and negative components influencing concept generation.

Demonstrates effective model editing through CAD-Erase and CAD-Amplify algorithms.

Validates the importance of both component types in generating and controlling concepts.

Abstract

Diffusion models have shown remarkable abilities in generating realistic and high-quality images from text prompts. However, a trained model remains largely black-box; little do we know about the roles of its components in exhibiting a concept such as objects or styles. Recent works employ causal tracing to localize knowledge-storing layers in generative models without showing how other layers contribute to the target concept. In this work, we approach diffusion models' interpretability problem from a more general perspective and pose a question: \textit{``How do model components work jointly to demonstrate knowledge?''}. To answer this question, we decompose diffusion models using component attribution, systematically unveiling the importance of each component (specifically the model parameter) in generating a concept. The proposed framework, called \textbf{C}omponent \textbf{A}ttribution for \textbf{D}iffusion Model (CAD), discovers the localization of concept-inducing (positive) components, while interestingly uncovers another type of components that contribute negatively to generating a concept, which is missing in the previous knowledge localization work. Based on this holistic understanding of diffusion models, we introduce two fast, inference-time model editing algorithms, CAD-Erase and CAD-Amplify; in particular, CAD-Erase enables erasure and CAD-Amplify allows amplification of a generated concept by ablating the positive and negative components, respectively, while retaining knowledge of other concepts. Extensive experimental results validate the significance of both positive and negative components pinpointed by our framework, demonstrating the potential of providing a complete view of interpreting generative models. Our code is available \href{https://github.com/mail-research/CAD-attribution4diffusion}{here}.