Read the original at HF Daily Papers
Researchers propose ALLOT, a training-free method that allocates block-sparse attention for video diffusion transformers by optimizing layer-step densities to improve output fidelity.
Carried by: HF Daily Papers. First seen: .