AIGVQA: A Unified Framework for Multi-Dimensional Quality Assessment of AI-Generated Video

Jiarui Wang, Juntong Wang, Xiaorong Zhu, Huiyu Duan, Guangtao Zhai, Xiongkuo Min; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 3414-3421

Recent advancements in large multimodal models (LMMs) have driven significant progress in both text-to-video (T2V) generation and video-to-text (V2T) interpretation tasks. However, AI-generated videos (AIGVs) still face challenges in terms of perceptual quality, aesthetic quality, temporal quality, and text-video alignment. To address this, a scalable and human-aligned automatic evaluation model for AIGVs is in urgent need. The VQualA 2025 GenAI-Bench Challenge, held in conjunction with ICCV 2025, provides a critical platform for assessing the perceptual quality and text-video alignment of AI-generated videos. In response to this challenge, we present AIGVQA, a LMM-based multi-dimensional AIGV Quality Assessment metric, employs an all-in-one architecture to assess videos from holistic and fine-grained perspectives. We apply our framework to both tracks of the GenAI-Bench challenge, utilizing the TaobaoVD-GC dataset. In Track I (Overall Quality Prediction), our model achieves a final score of 0.71. In Track II (Multi-Dimensional Quality Prediction), it achieves a final score of 0.76. Comprehensive experiments demonstrate the strong competitiveness and generalization capability of our AIGVQA framework for the complex task of AI-generated video quality assessment. The codes of AIGVQA are available at https://github.com/IntMeGroup/AIGVQA.