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

Chinese AI video-generation models have outperformed leading US rivals in visual quality, but several models lag behind in safety safeguards, according to researchers at the University of Hong Kong. ByteDance's Seedance 2.0 Pro topped the ranking of 12 leading AI video-generation models, scoring 3.84 out of 5 for visual quality, outperforming Google DeepMind's Veo 3.1, and OpenAI's Sora 2 Pro.

Seedance 2.0 Pro also led in visual consistency, physical realism, aesthetic quality, and instruction compliance, maintaining its lead when audio capabilities were factored in. The researchers believe ByteDance's success is due to its core business model focused on short video, giving it access to a vast volume of video training data.

Other Chinese models like Vidu Q3 Pro and Kuaishou Technology's Kling 3 also demonstrated competitive visual performance. However, the study revealed a gap between technical performance and content safety, as some high-performing models generated legally prohibited visual content when prompted with extreme violence or illegal behavior.

Seedance ranked sixth in blocking harmful content, Vidu seventh, and Kling ninth. The evaluation used 71 Chinese-language prompts over text-to-video and image-to-video generation, assessed by 20 human evaluators. Despite high technical performance, some models like MiniMax's Hailuo 2.3, which ranked lowest in visual quality, emerged as the safest.

The findings highlight that high technical performance and robust safety guardrails are not mutually exclusive. ByteDance and Kuaishou have released updated models since the assessment, while US tech giants have also responded with new releases. Safety remains a growing concern for developers in both regions.