Accelerating Spatio-Temporal Attention for Video Diffusion on TPUs

T1Covered by 1 source · first reported by Google — The Keyword / AI / Research / DeepMind / Developers at 30 Sept, 16:31 UTC

To address the quadratic latency bottleneck of self-attention in high-resolution video diffusion models, developers implemented Sparse VideoGen (SVG) to dynamically route attention heads to highly structured spatial or temporal sparse masks. Translating this algorithmic sparsity into physical…

Sources

• T1Accelerating Spatio-Temporal Attention for Video Diffusion on TPUsGoogle — The Keyword / AI / Research / DeepMind / Developers30 Sept, 16:31