Reproduce “M3-AGIQA: Multimodal, Multi-Round, Multi-Aspect AI-Generated Image Quality Assessment” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chuan Cui, Kejiang Chen, Zhihua Wei, Wen Shen, Weiming Zhang, Nenghai Yu
M3-AGIQA presents a framework for evaluating AI-generated image quality across perceptual, correspondence, and authenticity aspects, leveraging multimodal large language models in a multi-round conversational process with xLSTM for score prediction. This approach achieves state-of-the-art performance on multiple benchmarks by closely aligning with human judgments and providing interpretable evaluations.
Key insights
- Integrating a multi-round conversational approach with knowledge distillation from powerful online MLLMs significantly enhances the local model's ability to provide human-aligned, multi-aspect image quality assessments.
- Leveraging the full sequence of vocabulary logits from the MLLM's `lm_head` as features, rather than raw hidden states, and processing them with an xLSTM head, effectively captures nuanced multimodal understanding for robust quality prediction.
- The proposed framework offers strong interpretability by generating intermediate descriptive reasoning, allowing insights into *why* an image received a particular quality score.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.