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

A University of Hong Kong study has revealed that Chinese AI video-generation models have outperformed their US counterparts in visual quality, while falling short in content safety. ByteDance's Seedance 2.0 Pro topped the ranking with a visual quality score of 3.84, outperforming Google DeepMind's Veo (3.47) and OpenAI's Sora 2 Pro (3.27).

The Chinese model excelled in visual consistency, physical realism, aesthetic quality, and instruction compliance, and maintained its lead when audio capabilities were considered. Other Chinese models also showed competitive visual performance, with Vidu Q3 Pro scoring 3.49 in combined audio-video performance and Kuaishou Technology's Kling 3 tied with Sora for fourth place in visual-only performance.

Despite these advances, there is a notable gap between technical performance and content safety. While high-performing models like Seedance generated prohibited visual content in response to extreme violence or illegal behavior prompts, their safety filtering was found to be inadequate. MiniMax's Hailuo 2.3 emerged as the safest model tested, with a content safety score of 3.63.

The evaluation, conducted using 71 Chinese-language prompts, assessed model versions available in April, excluding newer releases. Both Chinese and US tech giants have been rapidly upgrading their models following the assessment, with ByteDance launching Seedance 2.5 in July and Kuaishou releasing Kling 4.0 this month. US giants Google and OpenAI have also responded with model updates and safety initiatives.

Safety remains a growing concern for developers on both sides of the Pacific, with ByteDance and OpenAI taking steps to address it.