Origins of Creativity in Attention-Based Diffusion Models

TL;DR

This paper explores how self-attention mechanisms in diffusion models influence the emergence of globally coherent and creative image generation, extending existing theories from CNN-based models to those with self-attention.

Contribution

It extends the theoretical understanding of diffusion models by analyzing the role of self-attention in producing globally consistent and creative images.

Findings

Self-attention induces globally image-consistent arrangements of features.

Empirical verification on a crafted dataset supports the theory.

Diffusion models with self-attention generate more coherent images than CNN-only models.

Abstract

As diffusion models have become the tool of choice for image generation and as the quality of the images continues to improve, the question of how `creativity' originates in diffusion has become increasingly important. The score matching perspective on diffusion has proven particularly fruitful for understanding how and why diffusion models generate images that remain plausible while differing significantly from their training images. In particular, as explained in (Kamb \& Ganguli, 2024) and others, e.g., (Ambrogioni, 2023), theory suggests that if our score matching were optimal, we would only be able to recover training samples through our diffusion process. However, as shown by Kamb \& Ganguli, (2024), in diffusion models where the score is parametrized by a simple CNN, the inductive biases of the CNN itself (translation equivariance and locality) allow the model to generate samples that globally do not match any training samples, but are rather patch-wise `mosaics'. Notably, however, this theory does not extend to describe the role of self-attention in this process. In this work, we take a preliminary step in this direction to extend this theory to the case of diffusion models whose score is parametrized by a CNN with a final self-attention layer. We show that our theory suggests that self-attention will induce a globally image-consistent arrangement of local features beyond the patch-level in generated samples, and we verify this behavior empirically on a carefully crafted dataset.