U-DiTs: Downsample Tokens in U-Shaped Diffusion Transformers

TL;DR

U-DiTs introduce a novel downsampling approach in diffusion transformers, leveraging U-shaped architecture to reduce computation while maintaining or improving image generation performance.

Contribution

The paper proposes U-DiT models that incorporate token downsampling in U-shaped diffusion transformers, significantly reducing computation and enhancing performance over existing DiT models.

Findings

U-DiT outperforms DiT-XL/2 with only 1/6 of the computation cost.

Token downsampling improves efficiency without sacrificing quality.

U-DiT demonstrates superior performance in latent-space image generation.

Abstract

Diffusion Transformers (DiTs) introduce the transformer architecture to diffusion tasks for latent-space image generation. With an isotropic architecture that chains a series of transformer blocks, DiTs demonstrate competitive performance and good scalability; but meanwhile, the abandonment of U-Net by DiTs and their following improvements is worth rethinking. To this end, we conduct a simple toy experiment by comparing a U-Net architectured DiT with an isotropic one. It turns out that the U-Net architecture only gain a slight advantage amid the U-Net inductive bias, indicating potential redundancies within the U-Net-style DiT. Inspired by the discovery that U-Net backbone features are low-frequency-dominated, we perform token downsampling on the query-key-value tuple for self-attention that bring further improvements despite a considerable amount of reduction in computation. Based on self-attention with downsampled tokens, we propose a series of U-shaped DiTs (U-DiTs) in the paper and conduct extensive experiments to demonstrate the extraordinary performance of U-DiT models. The proposed U-DiT could outperform DiT-XL/2 with only 1/6 of its computation cost. Codes are available at https://github.com/YuchuanTian/U-DiT.