Seedance leads video quality while Hailuo tops safety in HKU's benchmark

The model that produces the strongest video is not necessarily the one that best handles risky prompts. HKU Business School's latest benchmark puts ByteDance's Seedance 2.0 Pro first for generation quality, while MiniMax's Hailuo-2.3 leads the published safety results. The separation gives users a more useful picture than a single winner across all tasks.[1]

The Artificial Intelligence Evaluation Lab, led by Jack Jiang, released the findings on 26 October. The team tested text-to-video and text-and-image-to-video generation across visual continuity, physical plausibility, aesthetics, instruction alignment, audio quality and audiovisual matching. These are measures of output quality; the assessment does not establish which model generates video fastest.[1]

Two rankings answer different questions

Seedance led all six generation dimensions and also recorded the highest overall audiovisual result, 3.86. It placed sixth in the safety table. Hailuo led safety but was ninth in the university's overall generation ranking. Veo 3.1 was strong on both, illustrating that the results are more varied than a simple quality-versus-safety trade-off.[1]

The safety evaluation used prompts involving categories such as misinformation, illegal activity, violence, explicit material, discrimination and copyright-protected content. A model's score reflects its behaviour under that test design. It should not be read as a universal probability that every output from the service is safe.[1]

The public generation table includes 12 models, while the safety table contains 11 entries and omits a score for Sora. That absence leaves its safety result unreported in this release; it does not place the model at either end of the ranking.[1]

An image can make the task harder

For most tested models, adding a reference image reduced generation quality. An image supplies detail but also adds constraints: the model must preserve visual information while following the text and producing coherent movement. The benchmark shows that these combined tasks can remain difficult even when text-only results look convincing.[1]

Wan 2.6 illustrates the difference, scoring 3.41 for text-to-video and 2.75 with image input. The release describes that as a 66% drop, but the listed values show a decline of 0.66 points, or about 19.4% relative to 3.41. The point difference is the clearest way to report it without implying that a benchmark score is a linear measure of real-world quality.[1]

Choosing a model for a workflow

A production team may care about preserving a supplied character or product image, matching sound to movement, or reliably refusing a particular category of content. Those requirements call for different parts of the results. A visually impressive sample cannot answer all of them, and a missing audio capability should not be confused with a poor audio score.[1]

The benchmark is most useful as a set of questions to bring to a real task: does the model preserve the reference, follow the instruction, maintain continuity and behave appropriately under difficult prompts? HKU's results show why those questions should remain separate even when one model takes the top generation position.[1]

Report an error or suggest a correction