Beyond Source Prompts: Towards Visually Grounded AI-Generated Video Quality Assessment
VGCap is a framework for assessing AI-generated video quality using captions that are grounded in visual evidence rather than relying solely on source prompts. It creates two types of captions: a global caption refining the source prompt and segment captions describing localized motion. These complementary textual cues are fused with video features to evaluate spatial quality, temporal quality, and text-video alignment more accurately.