Recently, the TSAIL Lab at Tsinghua University has officially open-sourced the video generation acceleration framework TurboDiffusion in collaboration with Shengshu Technology. This breakthrough technology boosts the inference speed of AI video diffusion models by 100 to 200 times while maintaining nearly lossless visual quality, marking the official entry of AI video creation into the real-time generation era.

Remarkable Acceleration: From Minutes to Seconds

TurboDiffusion is deeply optimized for existing open-source video generation models (such as the Wan2.1 and Wan2.2 series of 1.3B to 14B parameter models). On a single RTX5090 GPU, the end-to-end generation speed can reach over 200 times the peak performance. For example:

- A 5-second video generation time was reduced from 184 seconds to 1.9 seconds (approximately 97 times faster);

- A high-resolution 720P video was reduced from thousands of seconds to tens of seconds on the 14B model.

Even on consumer-grade RTX4090 GPUs, it can run large models smoothly without relying on expensive A100 or H100 professional cards. This greatly lowers the hardware threshold for AI video generation, allowing more developers and creators to get started easily.

image.png

Core Technology Revealed: A Combination of Multiple Optimizations

TurboDiffusion is not a new model but an efficient inference framework that achieves near-lossless acceleration through systematic innovations. The main technologies include:

- 8-bit Quantization: Uses SageAttention for low-bit attention computation, accelerated by Tensor Core;

- Sparse Linear Attention (SLA): Retains only key attention connections, further reducing computational load;

- Time Step Distillation (rCM): Compresses traditional hundreds of sampling steps to 3-4 steps, combined with score regularization and continuous time consistency distillation to ensure stable quality.

These techniques are orthogonal and can be replicated with just 6 training steps, significantly improving operability. The project is fully open source, including code, pre-trained model weights, and complete training scripts, and has already gained widespread attention on GitHub.

D深远 Industry Impact: Opening a New Era of Real-Time AI Video

The release of TurboDiffusion is considered a "DeepSeek moment" in the field of video generation. It not only solves the latency and cost issues of high-resolution, long-sequence video generation but also attracted the attention of international teams such as Meta and OpenAI. This framework promotes the transition of AI video from experimental stages to large-scale applications, and is expected to support interactive creation and enterprise-level mass production in the future.