Tsinghua University and Shengshu Tech have made a groundbreaking leap in AI video generation with the open-sourcing of TurboDiffusion. This innovative framework revolutionizes the process, reducing video generation time from minutes to mere seconds on a single GPU. The impact is profound, potentially leveling the playing field for high-quality video creation, making it accessible to individual creators and small teams. Benchmark tests demonstrate remarkable speed improvements, showcasing a 97x acceleration for a 1.3B-parameter, 480p text-to-video model on an RTX 5090 GPU. The key to this success lies in SageAttention, a core component integrated into NVIDIA TensorRT, which has already found its way into products from industry giants like Tencent, ByteDance, and Alibaba. This development not only accelerates video generation but also opens up new possibilities for content creators, marking a significant step towards democratizing high-quality video production.
