[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"aihot-art-98984":3},{"itemId":4,"vertical":5,"category":6,"source":7,"score":8,"title":9,"summary":10,"analysis":11,"url":12,"coverUrl":13,"direction":13,"marketSignal":13,"publishedAt":14},"98984","ai","论文研究","Apple Machine Learning Research",62,"使用校准稀疏注意力加速文本到视频生成","苹果研究提出校准稀疏注意力方法，通过跳过低贡献token连接加速文本到视频生成。","苹果研究团队提出校准稀疏注意力方法，加速文本到视频生成模型。\n· 研究发现，时空注意力中大量token连接分数可忽略，且模式重复，可跳过计算而不影响结果。\n· 该方法适用于局部token块，显著减少计算量。\n· 基于扩散模型的视频生成因大Transformer骨干而速度慢，稀疏注意力直接解决这一瓶颈。\n影响\u002F看点：该技术有望大幅降低视频生成的计算成本，推动实时或近实时视频生成应用。","https:\u002F\u002Fmachinelearning.apple.com\u002Fresearch\u002Fcalibrated-sparse-attention",null,"2026-07-21 08:00:00"]