University of Hong Kong researchers scored 12 leading AI video generators on visual quality and on blocking harmful content. ByteDance’s Seedance topped quality; the safest model ranked near the bottom on looks.
2026-10-26 · by Moe Ameen
Researchers at the University of Hong Kong’s business school compared 12 leading AI video-generation models and scored each on two separate axes: visual quality and safety — how reliably the model blocked harmful content. The lead author is Zhenhui Jack Jiang, a professor of innovation and information management at HKU. The finding reported by the South China Morning Post on October 26, 2026 is that the two rankings barely overlap — the models that looked best were often not the ones that refused the riskiest prompts.
On visual quality, scored out of 5 by human evaluators, ByteDance’s Seedance 2.0 Pro ranked first at 3.84, ahead of Google DeepMind’s Veo 3.1 at 3.47 and OpenAI’s Sora 2 Pro at 3.27 — a result the coverage framed as Chinese models beating US rivals on looks. On safety the order flips: MiniMax’s Hailuo 2.3, which ranked near the bottom on visual quality at 2.95, came out the safest at 3.63, narrowly ahead of Veo at 3.62. Seedance landed mid-pack on blocking harmful content despite topping the quality chart.
The core safety finding is that several technically strong models generated legally prohibited imagery when prompted about extreme violence or illegal behaviour — exposing weaknesses in content filtering. The study used 71 Chinese-language prompts spanning text-to-video and image-to-video generation, with outputs judged by 20 human evaluators.
Treat the specifics as a single, relatively small snapshot. The evaluation covered model versions available as of April 2026 and excluded later releases — and the field has shipped several new versions since. The prompts were Chinese-language. Coverage also gives slightly different counts for the safety ranking (one summary places Seedance sixth of 11 on blocking harmful content), so treat exact placements loosely. The durable takeaway is the shape of the result: quality and safety are separate axes, and leading on one does not imply the other.
The study’s real message for a creator is that you should not bet your brand on one raw model’s twin gambles — that it looks good AND that its filter holds. [Kompozy](/) takes that bet off the table by putting a governed layer between generation and your feed. Instead of prompting a bare video model and publishing whatever comes back, you generate inside Kompozy against your [Persona Brief](/glossary/persona-brief) and banned-word filters, and nothing reaches a platform until it clears a per-post review in the pipeline. That human gate is exactly the safeguard the study found the models themselves cannot be trusted to provide.
On the capability side, Kompozy generates the video too — [Persona Shorts](/glossary/persona-shorts) (an avatar talking-head with auto-captions), [Clipped Shorts](/glossary/clipped-short) (a long clip cut to vertical with word-synced captions), Marketing Shorts, and Listicle Video — then fans the same idea into carousels, quote graphics, per-platform text, a blog, and a newsletter, all in one brand voice. [Autopilot](/glossary/autopilot) schedules and publishes the set across the eight social platforms plus blog and email. So whichever foundation model wins next quarter’s quality chart, your workflow does not change: Kompozy is the brand-safe, multi-platform layer that turns a generated clip into finished, captioned, scheduled posts — the part the leaderboard never measured. To weigh the raw tools yourself, see our pages on [Seedance](/ai-tools/seedance), the [Veo 3.1 API](/ai-tools/veo-3-1-api), and [Hailuo](/ai-tools/hailuo-ai-video-generator).
It depends on the axis. In the HKU study, ByteDance’s Seedance 2.0 Pro scored highest on visual quality (3.84/5), ahead of Veo 3.1 and Sora 2 Pro. On safety the ranking flipped: MiniMax’s Hailuo 2.3 scored highest (3.63) despite a near-bottom visual quality of 2.95. "Best" depends on whether you are optimizing for looks or for guardrails.
With a caveat. The study found several high-quality models would generate legally prohibited imagery on certain prompts, so filtering is uneven and the liability for what publishes falls on the creator or brand. The practical safeguard is a human review step between generation and publishing, which is why tools built for brands keep a per-post approval gate.
Treat it as directional. It is a single, relatively small study of 71 Chinese-language prompts judged by 20 evaluators, covering model versions available in April 2026 and excluding later releases; coverage even reports slightly different safety placements. The durable takeaway is that quality and safety are separate axes — not that any one model is permanently ranked.
Do not pick on demo reels alone. Quality has converged enough that several models make a usable clip, so the real differentiator is turning that clip into captioned, on-brand posts across platforms safely. Kompozy generates the video and governs it with a Persona Brief and a per-post review before publishing to eight social platforms plus blog and email.