Towards Stabilized and Efficient Diffusion Transformers through Long-Skip-Connections with Spectral Constraints

TL;DR

This paper introduces Skip-DiT, a diffusion transformer variant with Long-Skip-Connections that stabilize feature propagation, enabling faster training and inference while maintaining high output quality in image and video generation.

Contribution

It proposes Long-Skip-Connections for diffusion transformers, improving stability and efficiency, and provides theoretical and empirical validation of their benefits.

Findings

4.4x training acceleration

1.5-2x inference speedup

High fidelity with negligible quality loss

Abstract

Diffusion Transformers (DiT) have emerged as a powerful architecture for image and video generation, offering superior quality and scalability. However, their practical application suffers from inherent dynamic feature instability, leading to error amplification during cached inference. Through systematic analysis, we identify the absence of long-range feature preservation mechanisms as the root cause of unstable feature propagation and perturbation sensitivity. To this end, we propose Skip-DiT, an image and video generative DiT variant enhanced with Long-Skip-Connections (LSCs) - the key efficiency component in U-Nets. Theoretical spectral norm and visualization analysis demonstrate how LSCs stabilize feature dynamics. Skip-DiT architecture and its stabilized dynamic feature enable an efficient statical caching mechanism that reuses deep features across timesteps while updating shallow components. Extensive experiments across the image and video generation tasks demonstrate that Skip-DiT achieves: (1) 4.4 times training acceleration and faster convergence, (2) 1.5-2 times inference acceleration with negligible quality loss and high fidelity to the original output, outperforming existing DiT caching methods across various quantitative metrics. Our findings establish Long-Skip-Connections as critical architectural components for stable and efficient diffusion transformers. Codes are provided in the https://github.com/OpenSparseLLMs/Skip-DiT.