Recent advances in generative AI have made video generation increasingly capable, but AI-generated videos (AIGVs) still exhibit a range of distortions. A new preprint on arXiv, titled "Beyond Source Prompts: Towards Visually Grounded AI-Generated Video Quality Assessment," highlights this problem and argues that conventional quality metrics are insufficient for these outputs.

The authors propose moving beyond source prompts—the text or initial conditions used to generate a video—toward a visually grounded assessment framework. While the abstract does not detail the exact method, the emphasis is on evaluating the actual visual content of AIGVs rather than relying on the prompt as a reference.

This direction reflects a broader need in the field: as generative video becomes more common, dedicated quality assessment tools will be essential for benchmarking and improving models. The paper is a contribution to that effort, though its specific findings and experimental results are not described in the available abstract.