(Yicai) Dec. 5 -- Tencent Holdings has released a text-to-video generation tool based on the Chinese tech giant's Hunyuan artificial intelligence foundation model to rival OpenAI's Sora.

HunyuanVideo is the most parameter-rich and high-performing text-to-video model available in the open-source domain, Tencent announced on Dec. 3. With 13 billion parameters, it can generate five-second videos with high physical accuracy and scene consistency, turning concepts into reality and fostering creative expression, it added.

HunYuanVideo scored 41 percent when evaluated across various dimensions, including duration, text alignment, motion quality, and visual quality, according to the Tencent HunYuan team. It scored higher than two domestic tools and international models, such as Runway GEN-3 Alpha and Luma's Dream Machine 1.6, the team added.

Text-to-video technology is still immature, with most models having low success rates, Kai Sa, head of Tencent's HunYuan multimodal generation team, told Yicai. According to the Shenzhen-based company's internal evaluation, the tech is not yet at a level suitable for large-scale commercialization and still needs technical refinement, Kai added.

Unlike text-to-image models, which generate a single image each time, text-to-video models create 129 pictures per video, significantly increasing the computational demand, Kai noted. Many peers are reluctant to open-source such costly models, making them inaccessible to many, so HunYuan decided to open-source its video tool, Kai said.

Furthermore, video models struggle to simulate physical laws accurately, and the data processing, cleaning, and incorporation of physical laws involved are extremely complex, Kai pointed out, adding that HunYuan plans to incorporate real-world knowledge into HunyuanVideo.

Other Chinese companies have also launched video generation tools, including Kuaishou Technology's Keling AI, Tsinghua University-backed Shengshu AI's Vidu, Zhipu AI's Qingying, ByteDance's Dreamina AI, MiniMax's abab-video-1, and Alibaba Group Holding's Tongyi Qianwen. California-based Open AI's Sora is in its internal testing phase and unavailable to the public.

Editor: Martin Kadiev