Chinese video AI models beat US rivals in visual quality, but lag on safeguards: HKU study

The University of Hong Kong study, led by Zhenhui Jack Jiang, found ByteDance’s Seedance 2.0 Pro to be the leading AI video-generation model for visual quality, scoring 3.84 out of 5. This performance exceeded Google DeepMind’s Veo 3.1, which scored 3.47, and OpenAI’s Sora 2 Pro, which received 3.27. Seedance's lead extended across key visual aspects, including consistency in appearance, adherence to physical realism, aesthetic appeal, and its ability to comply with user instructions.

The research further indicated that Seedance 2.0 Pro maintained its superior standing even when audio capabilities were integrated into the evaluation. In this combined assessment, the ByteDance model scored 3.86, compared to Veo’s 3.51 and Sora’s 3.35. This suggests a robust overall performance that encompasses both visual and auditory elements in its generated content, which is crucial for producing comprehensive video content.

While demonstrating an edge in visual and audio quality, the study also identified a notable area where several Chinese AI video models lagged: safety safeguards. The report highlighted that while these models achieve high visual fidelity, some fall behind in mechanisms designed to block harmful content, with ByteDance’s Seedance ranking sixth out of eleven models for this specific criterion. This indicates a potential trade-off between advanced generation capabilities and content moderation.