A new independent ranking, and a surprising split
On October 26, 2026, the University of Hong Kong published a study that benchmarked 12 leading AI video-generation models on quality and safety, using 71 Chinese-language prompts and 20 human evaluators. It's one of the first independent academic comparisons of its kind, and the headline finding cuts against the usual narrative: Chinese models now lead on visual quality, while several models from both the US and China lag on safety guardrails.
The study, led by Zhenhui Jack Jiang, a professor of innovation and information management at HKU's business school, put ByteDance's Seedance 2.0 Pro at the top of the visual-quality rankings with a score of 3.84 out of 5. That's ahead of Google DeepMind's Veo 3.1 (3.47) and OpenAI's Sora 2 Pro (3.27). Seedance also led on visual consistency, physical realism, aesthetic quality, and instruction compliance, and it kept its lead when audio capabilities were factored in, scoring 3.86 against Veo's 3.51 and Sora's 3.35.
Jiang said Seedance's performance was unexpectedly strong, calling it 'far ahead' of both Chinese and American competitors. His explanation: ByteDance's core short-video business likely gave it access to a volume of training video that competitors simply don't have.
The best-looking model isn't the safest one
Here's the part that matters most if you're choosing a tool, not just admiring benchmarks: the model with the best visuals ranked only sixth out of 11 for blocking harmful content. According to the study, some models — the brief does not specify which — produced illegal content when prompted with extreme violence. The researchers were clear that safety and quality aren't mutually exclusive; they explicitly urged AI companies to close the gap rather than treating strong visuals as a substitute for safeguards.
Other Chinese models also showed up well on the quality side. Vidu Q3 Pro, from startup Shengshu Technology, scored 3.44 in visual-only performance and 3.49 combined with audio, placing third in both categories. The brief doesn't give us a full safety ranking across all 12 models, so it's not possible to say definitively which tools are safest overall — only that the study found a real, measurable gap between how good a model's output looks and how reliably it refuses unsafe prompts.
Why this matters even if you'd never type an unsafe prompt
If you're a careful creator, your first reaction might be: I don't prompt for extreme content, so why does this affect me? Two reasons worth taking seriously.
First, safety guardrails aren't just about deliberately extreme prompts. A model with weak content filters can misfire on ambiguous or borderline requests — action sequences, dramatic conflict, stylized violence for a thriller-style episode — in ways you didn't intend. If you're producing episodic content at volume, you're generating a lot of shots, and even a small per-generation risk of a policy-violating output compounds over a season's worth of episodes.
Second, the risk isn't just to the clip itself, it's to your distribution. Platforms like TikTok, Instagram Reels, and YouTube Shorts penalize accounts that post flagged or borderline content, sometimes with reduced reach, sometimes with strikes against the account. A single risky clip from a model with weak safeguards can do the same damage to your distribution that a bad hook or a visible watermark would — except it's harder to predict and harder to catch before you hit publish.
What creators should actually do with this information
The practical takeaway isn't to avoid Seedance or any specific model — the study doesn't say these tools are unsafe for ordinary use, only that their content guardrails are inconsistent relative to their visual quality. The takeaway is that safety track record deserves a spot in your tool-selection checklist, not just an afterthought you consider after something goes wrong.
- Treat independent benchmarks like this HKU study as a starting point for due diligence, not a final verdict — ask what a model's safety ranking was, not just its visual score.
- If you're producing episodic or serialized content at volume, build in a manual review step for anything involving conflict, violence, or mature themes, regardless of which model you're using.
- Keep a note of which platform guidelines apply to your niche. A model that technically generates a clip doesn't mean that clip is safe to post under TikTok's or Instagram's community standards.
- Don't assume audio-inclusive scores tell you anything about safety. Seedance's lead held even when audio was factored in (3.86 vs Veo's 3.51 and Sora's 3.35), but that's a quality metric, not a safety one — the two are tracked separately in this study for a reason.
The bigger picture for the AI video toolkit
This study is a useful reminder that the AI video landscape is moving fast enough that quality leaderboards change quickly, but safety infrastructure seems to be lagging behind across the board — not just for one company or one country. Jiang's team framed this as fixable: safety and quality can coexist, and the responsibility sits with the model developers to close the gap, not with creators to work around it.
In the meantime, if you're building a consistent cast and posting on a schedule, it's worth treating model selection as an ongoing decision rather than a one-time choice. Tools like Episodo that help creators manage episodic AI video production can make it easier to apply consistent review steps across a season, but no production workflow replaces checking what independent research actually says about the tool you're using before you scale up your output.