HKU's AI Video Benchmark: Looks vs Safety — 5 Takeaways for Creators
On October 26, 2026, HKU Business School's Artificial Intelligence Evaluation Lab (AIEL), led by Professor Jack Jiang, published a head-to-head of twelve AI video models. SCMP summed up the vibe: Chinese models beat US rivals on visual quality, with Seedance ahead of Veo and Sora. The safety table tells a different story.
1. Why care now?
About 430,000 micro-dramas shipped in Chinese Mainland in the first eight months of 2026, and AI-generated titles were more than 90% of new releases. Video models are production infrastructure.
2. Who won on looks?
Overall visual score out of 5: Seedance 2.0 Pro 3.84 (also #1 on all six dimensions; AV matching 3.86), Veo 3.1 3.47, Vidu Q3 Pro 3.44, Kling 3.0 3.27, Sora 2 Pro 3.27, Seedance 1.5 Pro 3.25, Wan 2.6 3.22, Kling O1 3.04, Hailuo-2.3 2.95, CogVideoX-3 2.92, PixVerse V5.5 2.82, Pika 2.2 2.81.
3. Who won on safety?
Hailuo-2.3 3.63, Veo 3.1 3.62, Pika 2.2 3.52, PixVerse V5.5 3.48, CogVideoX-3 3.35, Seedance 2.0 Pro 3.12 (#6), Vidu Q3 Pro 3.08, Kling O1 2.50, Kling 3.0 2.40, Wan 2.6 2.32, Seedance 1.5 Pro 2.18. Sora was not in the safety table — do not invent a score. Some models still produced inappropriate content on violence/illegal prompts.
4. The uncomfortable trade-off
Seedance 2.0 Pro is #1 on generation and #6 on safety. Hailuo is #1 on safety and roughly #9 on generation. Text+image often scored worse than text-only (Wan 2.6: 3.41 text-only vs 2.75 with image).
5. What should you actually do?
- Chasing look → Seedance 2.0 Pro leads this panel; still run human safety review.
- Brand, kids, regulated → prefer Hailuo or Veo even if the footage is less pretty.
- Don't assume a reference image helps.
- Expect safety to become the next competitive axis; researchers call for better A/V sync, image+text understanding, longer coherent scenes, and stronger guardrails on misinfo, copyright, and ethics.
Sources: HKU Business School / AIEL report, Professor Jack Jiang, SCMP coverage.
Comments
Post a Comment