Recently, Shengshu Technology and the TSAIL Lab at Tsinghua University jointly released a video generation acceleration framework called TurboDiffusion and open-sourced it. The release of this new framework has attracted widespread attention, with many people expecting it to bring breakthroughs in video generation technology. According to the official introduction, TurboDiffusion can achieve up to 200 times faster video generation inference without significantly affecting the generation quality.

image.png

The core technical advantages of TurboDiffusion lie in breaking through a key bottleneck in the field of video generation. Although previous video diffusion models have strong creative capabilities, their efficiency is limited due to high computational complexity, making them difficult to be widely applied. TurboDiffusion is not just a simple optimization solution, but rather a systematic combination of multiple cutting-edge technologies, achieving overall speed improvements from model computation, attention mechanisms to inference processes.

The framework adopts several innovative technologies to achieve acceleration. For example, the low-bit attention acceleration technology SageAttention can accelerate attention calculations without loss on low-bit Tensor Cores. The sparse-linear attention acceleration uses a trainable sparse attention Sparse-Linear Attention (SLA), which can achieve 17-20 times attention sparsity acceleration on top of SageAttention. In addition, TurboDiffusion also introduces the latest distillation method rCM, allowing the model to generate high-quality videos in only 3-4 steps, significantly improving the generation speed.

While maintaining high-quality output, TurboDiffusion has achieved a significant increase in video generation speed, making high-quality video generation gradually approach the feasible range of real-time interaction. This marks the entry of AI video creation into the "real-time generation" era, accelerating the industry's transition from the technology exploration phase to large-scale and commercial application stages.

TurboDiffusion: https://github.com/thu-ml/TurboDiffusion

Key points:  

📈 The TurboDiffusion framework achieves up to 200 times faster video generation while maintaining generation quality.  

🔍 It uses innovative technologies such as low-bit attention and sparse attention to improve video generation efficiency overall.  

🚀 The open-source nature of this framework provides new opportunities for research and application in the field of video generation, marking the entry into the "real-time generation" era.