Sparsity Should Be ALLOTted by Output: Training-Free Allocation of Block-Sparse Attention for Video Diffusion Transformers
ALLOT reports an 11.8 dB PSNR gain on HunyuanVideo at equal or shorter runtime with XAttention.
The researchers found that reducing attention computation affects final-video fidelity differently across layers and denoising steps, while the most sensitive pairs remain largely consistent across inputs. Their training-free method probes those pairs and calibrates their effects on the denoising trajectory, then allocates attention density within a measured latency budget. It replaces allocation while retaining existing block-selection strategies and kernels. Across four additional published pipelines, ALLOT outperformed or tied the default allocation in all nine tested settings.
HF Daily Papers' note
The researchers found that reducing attention computation affects final-video fidelity differently across layers and denoising steps, while the most sensitive pairs remain largely consistent across inputs. Their training-free method probes those pairs and calibrates their effects on the denoising trajectory, then allocates attention density within a measured latency budget. It replaces allocation while retaining existing block-selection strategies and kernels. Across four additional published pipelines, ALLOT outperformed or tied the default allocation in all nine tested settings.
HF Daily Papers' note
score 5