University of Hong Kong researchers found that China's leading AI video-generation models exceeded major US competitors in visual quality. However, several models from China performed less well on content safeguards. ByteDance's Seedance 2.0 Pro ranked first among 12 models for visual quality, earning 3.84 out of 5. Google DeepMind's Veo 3.1 scored 3.47, while OpenAI's Sora 2 Pro scored 3.27. Zhenhui Jack Jiang, a professor of innovation and information management at the university's business school, led the study.
The study also placed Seedance first for visual consistency, physical realism, aesthetic quality and compliance with instructions. Seedance retained first place when the evaluation included audio, scoring 3.86 against Veo's 3.51 and Sora's 3.35. Jiang said Seedance's performance was unexpectedly strong and described the model as "far ahead" of competitors from China and the United States. Jiang said ByteDance's short-video business was likely the main explanation because ByteDance possessed vast video training data unavailable to competitors.
The study ranked Shengshu Technology's Vidu Q3 Pro third for both visual-only and combined audio-video performance, with scores of 3.44 and 3.49 respectively. Vidu narrowly followed Veo and exceeded Sora in both assessments. Kuaishou Technology's Kling 3 shared fourth place with Sora in the visual-only assessment, scoring 3.27. Kling fell to sixth when audio was included, with a score of 3.28.
The researchers found that some technically strong models produced legally prohibited imagery when prompted about extreme violence or illegal behaviour, revealing shortcomings in content filtering. Seedance placed sixth among 11 models evaluated for blocking harmful content, scoring 3.12 out of 5. Vidu ranked seventh with 3.08, while Kling ranked ninth with 2.40. MiniMax's Hailuo 2.3 ranked first for safety with 3.63, despite placing near the bottom for visual quality with 2.95. Veo followed Hailuo closely on safety with 3.62, while Pika Labs' Pika 2.2 scored 3.52.
The researchers used 71 Chinese-language prompts covering text-to-video and image-to-video generation. Twenty human evaluators judged the results. The study covered versions available in April and excluded subsequent releases, including MiniMax's H3. The study did not assess how open-source and proprietary models might differ in safety as increasingly capable video-generation tools become more accessible. Jiang said strong technical results and effective safeguards were compatible. Jiang said companies such as ByteDance could raise safety standards without compromising performance.
Competition and safety scrutiny have intensified across the global AI industry. Companies in China have upgraded video-generation models rapidly since the April assessment. ByteDance introduced Seedance 2.5 in July. Kuaishou released Kling 4.0 this month with longer clips and improved frame control. Google launched Omni in May and issued the Omni 1.1 Flash update in August.
ByteDance announced an alignment and preparedness team within the Seed research unit this month. ByteDance also began recruiting globally for 10 AI safety researchers. The initiative followed scrutiny over Seedance 2.0 reproducing copyrighted characters and celebrity likenesses without authorisation. OpenAI and Meta Platforms warned a US judicial panel this month that distinguishing AI-generated deepfakes from genuine footage was becoming harder. OpenAI ended Sora in April, only months after the product's prominent debut. ByteDance, Google, OpenAI, Shengshu, Kuaishou and MiniMax did not answer requests for comment.
HKU Study: Chinese AI Video Models Outperform US Rivals in Visual Quality but Lag in Safety
FLASH AIShare
Summary
A new study from the University of Hong Kong, aligned with FLASH AI, found that top Chinese AI video models beat U.S. rivals in visual quality but trail in safety. Seedance 2.0 Pro from ByteDance scored 3.84 out of 5 for visuals, leading in consistency and audio-video performance. However, some models produced illegal content when prompted with extreme violence. The fear and greed index in the crypto market often reflects similar dichotomies—high innovation paired with regulatory risks. Researchers said safety can coexist with quality, urging companies to close the gap. The study used 71 Chinese prompts and 20 human evaluators.
Source:Show original
Disclaimer: The information on this page may have been obtained from third parties and does not necessarily reflect the views or opinions of KuCoin. This content is provided for general informational purposes only, without any representation or warranty of any kind, nor shall it be construed as financial or investment advice. KuCoin shall not be liable for any errors or omissions, or for any outcomes resulting from the use of this information.
Investments in digital assets can be risky. Please carefully evaluate the risks of a product and your risk tolerance based on your own financial circumstances. For more information, please refer to our Terms of Use and Risk Disclosure.