$\Delta$-DiT: A Training-Free Acceleration Method Tailored for Diffusion Transformers

TL;DR

This paper introduces $ ext{Delta}$-DiT, a training-free inference acceleration framework for diffusion transformers that leverages block-specific caching to significantly speed up image generation without sacrificing quality.

Contribution

The paper proposes a novel, training-free acceleration method for diffusion transformers, utilizing a cache mechanism tailored to DiT architecture and insights into block-function correlations.

Findings

Achieves 1.6x speedup on 20-step generation

Improves performance in most cases with acceleration

Outperforms existing methods at 4-step generation

Abstract

Diffusion models are widely recognized for generating high-quality and diverse images, but their poor real-time performance has led to numerous acceleration works, primarily focusing on UNet-based structures. With the more successful results achieved by diffusion transformers (DiT), there is still a lack of exploration regarding the impact of DiT structure on generation, as well as the absence of an acceleration framework tailored to the DiT architecture. To tackle these challenges, we conduct an investigation into the correlation between DiT blocks and image generation. Our findings reveal that the front blocks of DiT are associated with the outline of the generated images, while the rear blocks are linked to the details. Based on this insight, we propose an overall training-free inference acceleration framework $\Delta$-DiT: using a designed cache mechanism to accelerate the rear DiT blocks in the early sampling stages and the front DiT blocks in the later stages. Specifically, a DiT-specific cache mechanism called $\Delta$-Cache is proposed, which considers the inputs of the previous sampling image and reduces the bias in the inference. Extensive experiments on PIXART-$\alpha$ and DiT-XL demonstrate that the $\Delta$-DiT can achieve a $1.6\times$ speedup on the 20-step generation and even improves performance in most cases. In the scenario of 4-step consistent model generation and the more challenging $1.12\times$ acceleration, our method significantly outperforms existing methods. Our code will be publicly available.