TurboDiffusion - A video generation acceleration framework developed by Tsinghua University and others.

What is TurboDiffusion?

TurboDiffusion is a video generation acceleration framework jointly developed by Tsinghua University, BioNTech, and UC Berkeley. Through technologies such as SageAttention, SLA (Sparse Linear Attention), and rCM (Timestep Distillation), the framework increases video generation speed by 100-200 times, reducing a generation task from 184 seconds to 1.9 seconds on a single RTX 5090 graphics card. The framework lowers the barrier to entry for video generation, driving industry transformation and making creativity a core competitive advantage.

Main functions of TurboDiffusion

• Significantly accelerates video generationIt can increase video generation speed by 100-200 times. For example, on a single RTX 5090 graphics card, the generation task that originally took 184 seconds can be shortened to 1.9 seconds.
• High-quality video outputIt maintains video quality while accelerating the process, supporting the generation of 480p and 720p high-definition videos.
• Supports multiple modelsIt provides a variety of pre-trained models (such as Wan2.1 and Wan2.2) suitable for different tasks such as text-to-video (T2V) and image-to-video (I2V).
• Low resource demandBy quantifying and optimizing, the demand for hardware resources is reduced, making video generation more efficient and economical.

The technical principles of TurboDiffusion

• SageAttention (Low-bit attention mechanism)This approach replaces traditional attention mechanisms with low-bit SageAttention, reducing computational complexity while maintaining high-quality output. Combined with SLA (Sparse Linear Attention), it further optimizes long sequence processing and reduces computational burden.
• rCM (time-step distillation)By using time-step distillation technology, the multi-step diffusion process of traditional diffusion models is compressed into a few steps, significantly reducing the computational steps required for generation.
• Model quantization (W8A8)TurboDiffusion employs W8A8 quantization technology to quantize the model's weights and activation values to 8 bits. Quantization reduces the model's storage requirements and computational complexity, while improving inference efficiency and reducing GPU memory usage.
• Sparse activation and optimizationThis approach introduces a sparse activation strategy to reduce computation by selectively activating neurons. Combined with dynamic activation checkpointing, it further optimizes memory usage and computational efficiency, making it particularly suitable for high-dimensional data processing.

TurboDiffusion project address

• GitHub repositoryhttps://github.com/thu-ml/TurboDiffusion
• arXiv technical paper: https://arxiv.org/pdf/2512.16093

Application scenarios of TurboDiffusion

• Video content creationTurboDiffusion can quickly generate high-quality videos, suitable for advertising, film and television, short videos and other fields, significantly shortening the creation cycle and supporting real-time interactive scenarios.
• Film and television productionUsed for special effects production and video preview, it helps to quickly generate complex visual effects and script concept verification, reducing production time and costs.
• Advertising and MarketingThe framework can quickly generate personalized advertising videos to meet the needs of different scenarios, improve advertising appeal and production efficiency.
• Education and TrainingThe framework can generate educational videos and virtual training scenarios, enriching teaching content and improving learning outcomes and training experience.
• Games and EntertainmentGenerate dynamic videos in game development and interactive entertainment to enhance visual experience and immersion.