TraCache: Trajectory-Aware Feature Prediction for Training-Free Diffusion Transformer Acceleration

Chenxi Du, Yongheng Deng, Ju Ren, Yaoxue Zhang

Keywords: Diffusion Transformers, Inference Acceleration, Feature Caching

Abstract: Diffusion transformers have achieved remarkable success across various generative tasks but suffer from high inference costs. A promising line of work addresses this by reusing features across timesteps to minimize computational redundancy. However, existing methods degrade quality as temporal gaps increase due to trajectory shifts in the feature space. We propose TraCache, a trajectory-aware caching framework that models feature evolution across timesteps. Instead of direct reuse, TraCache fits local feature trajectories and extrapolates accurate predictions, mitigating drift while maintaining quality under aggressive acceleration. Extensive experiments on image and video generation show that TraCache significantly outperforms prior cache-based methods, especially in high skip-rate regimes. For instance, TraCache accelerates PixArt-$\alpha$ by 3.86$\times$ and Open-Sora by 3.74$\times$, while on DiT-XL/2, it provides a 4.51$\times$ acceleration with near-original visual fidelity.

Primary Area: generative models

Submission Number: 4358