HKU benchmark puts Seedance 2.0 Pro first on quality, Hailuo and Veo first on safety

HKU Business School’s AI Evaluation Lab, led by Professor Jack Jiang, has published a report scoring 12 AI video models. The lab tested text-to-video and image-to-video on six measures, from physical plausibility to how well the audio matches the picture.

ByteDance’s Seedance 2.0 Pro led on quality with 3.84 out of 5. Veo 3.1 came second (3.47) and Vidu Q3 Pro third (3.44). Kling 3.0 and Sora 2 Pro tied for fourth. PixVerse V5.5 and Pika 2.2 took the last two places. HKU also found that most models lost quality when given a starting image instead of text alone.

The safety ranking flips the order. HKU tested whether each model would make misinformation, violent or explicit content, biased content or copyrighted material. Hailuo 2.3 was safest (3.63), just ahead of Veo 3.1 (3.62) and Pika 2.2 (3.52). Seedance 2.0 Pro landed sixth of 11. Both Kling models, Wan 2.6 and Seedance 1.5 Pro made up the bottom four.

Keep in mind that the scores cover the versions HKU tested, not newer ones like Kling 4.0. If you make brand or client work, Veo 3.1 was the one model near the top of both lists.