PISA: Piecewise Sparse Attention Is Wiser for Efficient Diffusion Transformers

Name: PISA: Piecewise Sparse Attention Is Wiser for Efficient Diffusion Transformers
Uploaded: 2026-10-22T02:44:00.763748+00:00
Duration: 596 s
Description: PISA proposes Piecewise Sparse Attention to speed up Diffusion Transformers by using sub-quadratic complexity. It combines exact computation for critical attention blocks with block-wise Taylor expansion for others, creating a training-free method that maintains high quality and preserves pre-trained weights.

PISA proposes Piecewise Sparse Attention to speed up Diffusion Transformers by using sub-quadratic complexity. It combines exact computation for critical attention blocks with block-wise Taylor expansion for others, creating a training-free method that maintains high quality and preserves pre-trained weights.

Open the full episode