Copyright scmp
Chinese artificial intelligence start-up Shengshu AI on Tuesday rolled out Vidu Q2, a new video-generation model aimed at challenging OpenAI’s Sora, intensifying the global race to create realistic, controllable AI-generated videos. The model allows creators to produce expressive and consistent videos using up to seven reference images to define faces, gestures, scenes or props. ShengShu said the “Vidu Q2 Reference-to-Video” feature can merge multiple unrelated elements – such as distinct characters, objects and backgrounds – into a single coherent clip, improving realism, cinematic motion and prompt comprehension. “Vidu Q2 ‘Reference-to-Video’ marks a new chapter in AI video creation,” said Luo Yihang, CEO of Shengshu Technology. A former head of AI solutions and product lines at ByteDance’s cloud unit Volcano Engine, Luo said the model “goes beyond basic video creation; it’s about teaching AI to act and tell stories alongside creators”. The Vidu Q2, along with the model’s application programming interface for enterprises, is available globally through Shengshu’s platform. It supports reference-to-video, image-to-video and text-to-video generation, producing clips of up to eight seconds in various aspect ratios, and also features synchronised dialogue and sound effects. Membership tiers range from free access to a flagship plan costing up to 600 yuan (US$84) a month, which offers faster output and more generation capacity. Founded in March 2023, the Beijing-based company focuses on multimodal AI models through its Vidu video platform, which includes the previously released Vidu 1.5, Vidu 2.0 and Vidu Q1 versions. The firm’s leadership team has strong ties to Tsinghua University, including co-founder and now president Tang Jiayu, a Tsinghua graduate; Tsinghua University professor and company chief scientist Zhu Jun, and his student Bao Fan, who is the company’s chief technology officer. Shengshu has announced a slew of collaborations this year with partners such as the Beijing Film Academy, PC maker Lenovo, as well as a major online fiction platform to test AI-assisted content production. Shengshu said its user base exceeded 30 million in more than 200 countries, with annual recurring revenues crossing US$20 million just eight months after launch. The start-up recently raised hundreds of millions of yuan in series A funding led by Bohua Capital and including existing investors such as Baidu, Qiming Venture Partners, C&D Emerging Investment and the Beijing municipal government-backed AI Industry Fund. Previous backers include Huawei-linked Habo Investment, ByteDance-backed Jinqiu Capital, Alibaba Group Holding-affiliated fintech giant Ant Group and Zhipu AI. Alibaba owns the Post. The latest release comes as AI companies worldwide race to build models that better understand and simulate the physical world. OpenAI introduced Sora 2 in late September, describing it as “more physically accurate, realistic and controllable” than earlier systems. Google DeepMind’s Genie 3, unveiled in August, aims to serve as a “general-purpose world model” capable of producing diverse interactive environments. Earlier this month, Google updated its Veo 3.1 video generator. In China, the crowded field includes entrants such as short video firm Kuaishou’s Kling, ByteDance’s Jimeng, Alibaba’s open-sourced Wan under its Qwen series, and Tencent’s Hunyuan series model.
