Vidu Launches Globally, with Baidu’s AIHC Support for Large-Scale Video Model Training

On July 30th, Shengshu AI officially launched Vidu (www.vidu.studio), China’s first domestically developed large-scale video model and the world’s first challenger to Sora. The platform doesn’t require an application process, allowing users to start experiencing it right away by simply registering with their email.

Shengshu AI is one of the pioneers in China in the deployment of multimodal general-purpose large models. In April of this year, in collaboration with Tsinghua University, they introduced Vidu, a video model that stands toe-to-toe with OpenAI’s Sora. Since its initial unveiling in late April, Vidu has quickly garnered widespread attention both at home and abroad, thanks to its impressive text-to-long-video capabilities.

Vidu now fully supports text-to-video and image-to-video features, offering a choice of 4s and 8s durations with a maximum resolution of 1080P. The model continues to deliver high dynamism, realism, and consistency, as demonstrated in its April showcase. Furthermore, it has introduced new features such as character consistency (Character To Video), anime style, and text and special effect image generation.

In terms of user experience, Vidu has set a new industry standard with its rapid inference speed, generating a 4-second clip in just 30 seconds. This significantly outperforms mainstream AI video tools on the market, which typically require users to wait 1 to 5 minutes, or even longer, to generate a video clip of around 4 seconds. Vidu’s efficiency not only ensures a near-seamless creative experience, but it also allows the platform to easily meet the demands of a large user base, offering personalized services.

The rapid development of Vidu, from launch to going live within a mere three months, can be attributed to the engineering expertise of the Vidu R&D team, particularly their experience in efficient model training. The intensive and fast-paced model training tasks demand not only substantial computing power, but also a comprehensive approach that ensures quality from data preparation to model training and inference.

To facilitate model training, Vidu leverages Baidu‘s AI Heterogeneous Computing (AIHC) platform. Built around a GPU core, AIHC is a product of Baidu Intelligent Cloud’s years of AI experience and is particularly suited for large model training, ensuring both stability and efficiency.

AIHC excels at improving model training efficiency by managing computing resources across different locations, scales, and clusters. Its suite of technologies, including chip performance optimization, automatic chip selection, and tidal co-location, significantly enhance intelligent computing power efficiency. The platform achieves an effective training time ratio of over 98.8% on a ten-thousand-card cluster, with linear acceleration ratio and bandwidth effectiveness both exceeding 95%.

In addition to efficiency, Baidu‘s AIHC also focuses on the stability of model training. It offers a wealth of operation and observability tools and fault tolerance capabilities. These features enhance the stability of long-term operation of large-scale clusters, reduce the likelihood of training and inference tasks being prematurely terminated due to faults, and consequently minimize business losses.

SEE ALSO: AI Image Generation Platform LiblibAI Secures Multi-Billion Yuan Financing within a Year