Tsinghua University and Shengshu Tech Open-Source TurboDiffusion, Achieving “Second-Level” Video Generation on a Single GPU
Want to read in a language you're more familiar with?
Researchers from Tsinghua University and Shengshu Tech open-sourced TurboDiffusion, a framework that slashes AI video generation time from minutes to seconds on a single GPU, potentially democratizing high-quality video creation.
Tsinghua University’s TSAIL Lab, in collaboration with Shengshu Technology, has open-sourced TurboDiffusion, a new acceleration framework for AI video generation.
Benchmark tests show that on a single RTX 5090 GPU, a 1.3B-parameter, 480p text-to-video model can be accelerated from approximately 184 seconds to just 1.9 seconds, representing a 97× speedup.
The breakthrough shifts high-quality video generation to a “seconds-per-video” tool accessible to individual creators. Notably, its core component, SageAttention, has been integrated into NVIDIA TensorRT and deployed in products from Tencent, ByteDance, and Alibaba.
Source：liangziwei
