Rethinking Token-wise Feature Caching: Accelerating Diffusion Transformers with Dual Feature Caching
The episode discusses a paper titled "Rethinking Token-wise Feature Caching" which introduces Dual Feature Caching (DuCa) to accelerate Diffusion Transformers. The hosts conclude that alternating between aggressive and conservative caching, combined with random token selection, provides superior speed and quality compared to traditional importance-based methods.