BRACE: Barycentric Rational Forecasting for Fast Diffusion Transformers
A new method called BRACE, which stands for Barycentric Rational Forecasting with Chebyshev Enhancement, has been developed to speed up inference in Diffusion Transformers (DiTs) used for creating top-notch images and videos. Since DiTs require a lot of computational power, they've added temporal feature caching to cut down on unnecessary calculations. However, existing cache-then-forecast methods that use derivative-based polynomials can lose quality during fast processing, leading to unpredictable long-term predictions. BRACE addresses this issue by shifting from derivative-based methods to feature-driven rational forecasting. It uses a local sliding window to store selected historical features and applies modified Chebyshev weights for better accuracy. You can find the full details in a paper on arXiv with ID 2608.07572.
Key facts
- BRACE is a new method for fast inference in Diffusion Transformers.
- It uses barycentric rational forecasting with Chebyshev enhancement.
- It addresses quality degradation in existing cache-then-forecast methods.
- The method is motivated by the observation of sharp irregularities in DiT feature trajectories.
- It maintains a local sliding window to cache sparse historical features.
- It uses adapted Chebyshev weights for forecasting.
- The paper is available on arXiv with ID 2608.07572.
- The method targets high acceleration scenarios in image and video generation.
Entities
Institutions
- arXiv