VQQA: An Agentic Approach for Video Evaluation and Quality Improvement

A team of researchers has introduced VQQA (Video Quality Question Answering), a novel framework that tackles a core challenge in AI video generation: aligning outputs with complex user intent. Existing methods for improving video quality are often computationally expensive or require intrusive access to a model's internal workings. VQQA offers a unified, agentic approach that works across different input types—like text-to-video (T2V) and image-to-video (I2V)—by using a multi-agent system to dynamically generate visual questions. The answers from a Vision-Language Model (VLM) act as 'semantic gradients,' providing human-readable, actionable feedback instead of opaque numerical scores.

This feedback drives a highly efficient, closed-loop optimization process that operates through a simple natural language interface, treating the video generator as a black box. In extensive testing, VQQA proved exceptionally effective at isolating and resolving visual artifacts. It significantly boosted generation quality in just a few refinement cycles, achieving absolute improvements of +11.57% on the T2V-CompBench benchmark and +8.43% on VBench2 compared to standard 'vanilla' generation. These results substantially outperform current state-of-the-art techniques like stochastic search and other prompt optimization methods, demonstrating a more scalable and interpretable path to high-quality AI video.

The framework's agentic design allows it to generalize across tasks without retraining, making it a versatile tool for developers. By replacing passive evaluation metrics with an active critique-and-refine loop, VQQA shifts the paradigm from mere measurement to direct quality enhancement, paving the way for more reliable and controllable video generation models in practical applications.