Achieve faster generation, lower costs, improved user experience, and scalable enterprise use without compromising visual quality.
Singapore, December 23, 2025 /PRNewswire/ — ShengShu Technology and Tsinghua University TSAIL Lab jointly 100-200 times Faster Generate AI videos with little to no loss of visual quality. This release is a significant milestone for the industry, marking the arrival of the real-time generation era of AI video creation, or the video foundation model's “DeepSeek moment.”
Video generation is reaching a critical tipping point as generative AI advances rapidly in content creation. The focus is no longer just on being able to produce video, but on producing high-quality output faster, at lower cost, and at scale for real-world and enterprise use. To address the long-standing trade-off between quality, speed, and computing cost in high-resolution long-form video generation, ShengShu Technology and the Tsinghua University TSAIL Laboratory conducted fundamental research on inference efficiency, leading to the development of TurboDiffusion, designed to improve the practicality and scalability of AI video creation.
Since its release, TurboDiffusion has sparked widespread discussion across the international AI research and developer community, attracting attention from Meta and OpenAI researchers as well as teams behind major open source inference acceleration projects such as vLLM.
Breaking the speed barrier in high-quality video generation
Before the release of TurboDiffusion, ShengShu Technology had already established a strong position in AI video generation. In September 2024, Vidu introduced the world's first thematic consistency feature, ushering in a new era of reference-based video generation and gaining widespread adoption among creators.
The recent release of Vidu Q2 brings several more industry-leading features.
- A complete image generation stack covering text-to-image conversion, enhanced image references, and comprehensive image editing.
- Upgraded reference-based video generation with improved semantic understanding, camera control, and multi-subject consistency.
- Highly efficient image generation generates 1080p images in 5 seconds without compromising visual quality.
These results demonstrated that the benefits of Vidu are not achieved by sacrificing visual quality, but through a mature model architecture and strong engineering capabilities.
As video generation moves to higher resolutions, longer durations, and more complex application scenarios, the entire industry continues to face challenges related to latency and cost. TurboDiffusion was specifically developed to overcome these limitations.
Researchers and industry observers point out that TurboDiffusion's core technical advantages mark a watershed moment in video production. Diffusion-based video models have demonstrated strong creative potential but have long been constrained by computational complexity and efficiency limitations. TurboDiffusion achieves high-quality video generation within a practical range of near real-time interactions by significantly reducing generation latency while maintaining visual quality.
As a result, TurboDiffusion is widely recognized as a “DeepSeek moment” for video foundation models, accelerating the transition from experimental research to scalable real-world and commercial deployments, and marking the transition to real-time interactive AI video creation.
TurboDiffusion is not based on a single optimization. Instead, it systematically combines multiple advanced acceleration techniques to improve efficiency.
- Low bit attention acceleration
- Uses of TurboDiffusion Sage caution Perform attention computations on low-bit Tensor Cores and achieve lossless speedups of several times.
- Sparse linear attention acceleration
- TurboDiffusion employs trainable sparse attention; Sparse Linear Attention (SLA)sparse the calculation of attention and add additional 17～20× Speed up sparse attention in addition to SageAttend.
- Distillation acceleration of sampling step
- Using the most advanced distillation method rCMthis model can produce high quality videos: 3-4 steps.
- linear layer acceleration
- TurboDiffusion quantizes both the linear layer weights and activations to 8 bits (W8A8), which speeds up linear calculations and significantly reduces VRAM usage.
Combining these techniques enables near-lossless acceleration, allowing TurboDiffusion to deliver dramatic speed increases while maintaining visual stability and consistency.
These four core technologiesThe technology was independently developed by the Tsinghua University TSAIL team and ShengShu Technology. they carry Significance and far-reaching impact of milestones Contribute to both the breakthrough of AI multimodal foundation models and their industrial-scale deployment. especially, Sage caution is the first method to enable low-bit attention acceleration. It is already being implemented at scale across the industry.
for example, Sage caution has been successfully integrated into NVIDIA's inference engine TensorRT and is also deployed and operational on leading GPU platforms such as Huawei Ascend and Moore Threads S6000. Moreover, leading global and domestic technology companies and teams such as Tencent Hunyuan, ByteDance Doubao, Alibaba Tora, Shengshu Vidu, Zhipu Qingying, Baidu PaddlePaddle, Kunlun Wanwei, Google Veo3, SenseTime, vLLM, etc. have adopted this technology in their core products. Real economic value.
convert minutes to seconds
The impact of TurboDiffusion is huge. In the open source video generation model 1.3B/14B-T2V, turbo body fusion achieve From 100 times to the peak of 200 times End-to-end acceleration Single RTX 5090 GPUwhich significantly reduces generation time while maintaining visual quality. The code and models are open source and can be directly deployed.
Similar improvements are observed when applied to ShengShu Technology's proprietary Vidu video model. For example, producing a 1080p, 8-second high-quality video that used to take about 900 seconds now takes about 8 seconds. What once took minutes now takes seconds.
This transition brings AI video generation closer to real-time interaction, greatly increasing ease of use for both creators and businesses.
Going forward, ShengShu Technology will continue to invest in fundamental innovations to increase efficiency, improve user experience, and reduce creation and deployment costs. Through continued advances at the system and model level, the company aims to accelerate the real-world adoption of generative AI and usher the creative ecosystem into a new era of greater efficiency.
For more information:
TurboDiffusion: https://github.com/thu-ml/TurboDiffusion
SageAttend: https://github.com/thu-ml/SageAttend
Sparse Linear Attention: https://github.com/thu-ml/SLA
rCM: https://github.com/NVlabs/rcm
For more information about Vidu, please visit www.vidu.com.
The Vidu API is available at platform.vidu.com.
About ShengShu Technology
Founded in March 2023, ShengShu Technology is a world-leading artificial intelligence company specializing in the development of multimodal large-scale language models. The company delivers cutting-edge MaaS and SaaS products that drive innovation and revolutionize creative production by enabling smarter, faster, and scalable content creation. With its flagship video generation platform Vidu, ShengShu Technology's solutions span fields such as interactive entertainment, advertising, film, animation, and cultural tourism, and are delivered to more than 200 countries and regions around the world.
Source ShengShu Technology
