TurboDiffusion - Raw Digital Technology, Tsinghua and other open source video generation acceleration framework

What is TurboDiffusion?

TurboDiffusion is a video generation acceleration framework jointly open-sourced by Tsinghua University, BioDigital Technology and UC Berkeley, which is capable of increasing video generation speed by 100-200 times while maintaining nearly lossless picture quality. Through innovative means such as sparse linear attention, sample step distillation, and 8-bit quantization, the framework achieves the generation of 5-second HD video in 1.8 seconds on a single RTX 5090 graphics card. The core breakthrough lies in reducing the quadratic computational complexity of traditional video diffusion model to linear, solving the long-standing efficiency bottleneck in the industry, which has been paid attention by Meta, OpenAI and other organizations, and is known as the "DeepSeek Moment" of the large video model. It has been open-sourced and applied to multimodal large models such as Vidu.

Features of TurboDiffusion

• Low Bit Attention Acceleration: Using SageAttention technology, the attention computation is executed on a low-bit Tensor Core to achieve lossless and multi-fold acceleration.
• Sparse - Linear Attention Acceleration: Achieve a further 17 - 20x sparse acceleration of attention on top of SageAttention using trainable Sparse-Linear Attention (SLA).
• Sampling step distillation acceleration: Through the advanced distillation method rCM, the model is able to generate high quality video in only 3 - 4 steps.
• Linear Layer Acceleration: 8-bit 128x128 block granularity quantization (W8A8) of parameters and activations in the linear layer of the model, accelerating linear calculations and significantly reducing graphics memory usage.

Core Benefits of TurboDiffusion

• Efficient Reasoning: TurboDiffusion's inference is so fast that, for example, it takes only 1.9 seconds to generate a 16-frame video at 720P resolution on the RTX 5090.
• Maintaining high quality output: The quality of the generated video remains at a high level while speeding up dramatically.
• Open Source and Ease of Use: TurboDiffusion's code and model parameters (Checkpoints) are open source and can be installed and used by users in a few simple steps.

What is TurboDiffusion's official website

• GitHub repository:: https://github.com/thu-ml/TurboDiffusion
• arXiv Technical Paper:: https://arxiv.org/pdf/2512.16093

Who is TurboDiffusion for?

• creative worker: such as ad designers, video creators, animators, etc., can quickly generate creative videos to help them iterate and optimize their ideas quickly.
• research worker: Provide scholars and scientists working on video generation technologies with efficient experimental tools to accelerate the research process.
• schoolchildren: For students studying computer vision and artificial intelligence, TurboDiffusion can help them quickly understand and apply video generation techniques.
• Enterprise DevelopersTurboDiffusion can significantly improve content production efficiency in business applications that require rapid video content generation, such as short-form video platforms, social media, and online education.
• Individual enthusiasts: Individual users interested in video generation technology can use TurboDiffusion to quickly generate interesting videos for their personal entertainment and creative needs.