DPC-VQA: Decoupling Quality Perception and Residual Calibration for Video Quality Assessment
Researchers have developed DPC-VQA, a new framework for video quality assessment that leverages multimodal large language models (MLLMs). This approach decouples the perceptual capabilities of a frozen MLLM from a lightweight calibration branch, allowing for efficient adaptation to new scenarios without extensive retraining. DPC-VQA demonstrates competitive performance on both user-generated and AI-generated content benchmarks while significantly reducing trainable parameters and the need for MOS labels. AI