The headline result is clear enough. In a University of Hong Kong comparison of 12 leading video-generation models, ByteDance's Seedance 2.0 Pro came first for visual quality, scoring 3.84 out of 5. Google DeepMind's Veo 3.1 followed at 3.47. Shengshu Technology's Vidu Q3 Pro scored 3.44, ahead of OpenAI's Sora 2 Pro at 3.27, a score it shared with Kuaishou's Kling 3.
Seedance also led the individual visual measures the study reported: consistency, physical realism, aesthetic quality and following instructions. Adding audio did not change the order at the top: Seedance scored 3.86, Veo 3.51, Vidu 3.49 and Sora 3.35. Including audio raised Kling 3's score slightly, from 3.27 to 3.28, but pushed it down to sixth place as competing models gained more.
The study evaluated 12 models for quality but reported safety scores for 11. The available reporting does not identify which model was omitted from the safety comparison. The safety ranking points in a different direction. Among the 11 models assessed on this evaluation's harmful-content prompts, MiniMax's Hailuo 2.3 came first, at 3.63 out of 5. Veo was close behind at 3.62 and Pika Labs' Pika 2.2 scored 3.52. Seedance placed sixth, at 3.12. Vidu was seventh at 3.08 and Kling ninth at 2.40. According to the reporting, some models generated imagery the researchers classified as prohibited when prompted with extreme violence or illegal-behaviour scenarios.
Hailuo's position captures the split in one line. The model that blocked harmful prompts best scored 2.95 for visual quality, near the bottom of the 12. That does not make it a bad product, and it does not prove a necessary trade-off. It means a single ranking cannot answer two separate questions: which model looks best, and which model refuses what it should refuse.
The date is part of the result
The study used 71 Chinese-language prompts and 20 human evaluators, across text-to-video and image-to-video tasks. Those are reasonable choices for a comparative snapshot, and they are also limits. Because the prompts were in Chinese, the results may not generalize to English-language prompts or to other user populations. And 20 evaluators judging short outputs cannot settle a field that changes every few months.
The larger limit is timing. The versions assessed were available in April. MiniMax's H3 is explicitly outside the study. So are the releases that followed: ByteDance introduced Seedance 2.5 in July, Google launched Omni in May and updated it with Omni 1.1 Flash in August, and Kuaishou released Kling 4.0 this month. None of those newer releases was measured by this study. A reader choosing a video model today is choosing among a different set of products than the one ranked here.
That does not make the study useless. It makes it historical evidence about an April field, with a durable observation inside it: quality leadership and safety leadership sat with different companies, and neither score was close to the 5-point ceiling. Seedance's 3.84 and Hailuo's 3.63 both leave room that later models may already have taken, or lost.
Jiang, who led the study, pointed to ByteDance's short-video business as a likely explanation for Seedance's lead, citing video data that rivals cannot easily match. Jiang argues that visual quality and safeguards can coexist. The reported rankings do not disprove that possibility, but they do not demonstrate it either: the quality leader and safety leader were different models.
For anyone citing this study in a pitch, a procurement note or a post, the honest citation includes the month. Seedance 2.0 Pro led an April ranking. Hailuo 2.3 led its safety test. Neither sentence, on its own, describes the market in late October.
These figures come from October 26 reporting on the HKU study. The original study and complete scoring tables were not independently reviewed for this article.
Sources
- [1] South China Morning Post reporting via KuCoin Flash (October 26, 2026) — all scores, the method (71 prompts, 20 evaluators), the April scope and exclusion of H3, Jiang's comments, and the post-April releases (Seedance 2.5, Kling 4.0, Omni)Read source
- [2] Cybernoz (October 26, 2026), a second syndication of the same SCMP reporting — the headline scores (Seedance 3.84, Veo 3.47, Sora 3.27) and Jiang's role at HKU Business SchoolRead source