NVIDIA发的新论文，用时空缓存压缩和稀疏注意力，让视频扩散生成速度飙5-10倍，内存还不涨，搞视频生成和世界模型的朋友可以看看。
NVIDIA研究团队在ICML2026发表论文“Fast Autoregressive Video Diffusion & World Models with Temporal Cache Compression & Sparse Attention”。该方法通过时空缓存压缩和稀疏注意力解决自回归视频扩散中的注意力瓶颈。实验显示可实现5倍至10倍的推理速度提升，并在长序列生成中保持恒定内存占用。该工作同时适用于视频扩散模型和世界模型。
The NVIDIA research team ships papers faster than we can post about them. Here's another great pap...
The NVIDIA research team ships papers faster than we can post about them. Here's another great paper from #ICML2026 . Dvir Samuel @dvir_samuel 🚀 Excited to share our new paper: “Fast Autoregressive Video Diffusion & World Models with Temporal Cache Compression & Sparse Attention.” We address attention bottlenecks in auto-regressive video diffusion, enabling ×5–×10 speedup and constant memory over long rollouts. Your browser does not support the video tag. 🔗 View on Twitter 🔗 View Quoted Tweet 💬 6 🔄 3 ❤️ 51 👀 7312 📊 9 ⚡