HKU's video-model study finds different leaders for generation quality and safety
Seedance 2.0 Pro leads the published generation comparison, while Hailuo 2.3 leads a separate safety assessment under the study's tested conditions.
A University of Hong Kong study separates two questions that AI-video rankings can blur: how well a model generates a requested scene and how it handles high-risk requests. HKU Business School's October 26 announcement reports different leaders in the two assessments.
The school's Artificial Intelligence Evaluation Lab evaluated text-to-video and text-and-image-to-video generation. Its quality dimensions covered visual continuity, physical plausibility, aesthetics, instruction following, audio quality and alignment between audio and visuals. Separate safety prompts addressed areas including misinformation, illegal content, violence, discrimination and protected material.
In the published comparison, ByteDance's Seedance 2.0 Pro ranked first overall for generation and across all six quality dimensions, but sixth for safety and responsibility. Hailuo 2.3 led the safety assessment while placing ninth for generation. Those placements describe the tested models and conditions, rather than a permanent ranking of every current video service.
The researchers also report that adding a reference image reduced generation quality for most tested models. A reference can constrain appearance while adding another requirement the model must reconcile with the text prompt. That makes image-guided generation a separate capability to measure rather than an automatic improvement over text alone.
The study's useful contribution is the separation of these outcomes. A production team selecting a video tool may care about consistent movement and instruction following, while a platform must also consider how the system responds to harmful or misleading requests. A single overall score would hide the fact that success on one assessment did not imply leadership on the other.
Sources
- [HKU Business School Releases Report on Gen AI Video Generation Capabilities](https://lead.hku.hk/en/article/2026/10/hku-business-school-releases-report-on-gen-ai-video-generation-capabilities/) — University of Hong Kong