Reproduce “AesBench: An Expert Benchmark for Multimodal Large Language Models on Image Aesthetics Perception” 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
Yipo Huang, Quan Yuan, Xiangfei Sheng, Zhichao Yang, Haoning Wu, Pengfei Chen, Yuzhe Yang, Leida Li and 1 more
AesBench introduces the first expert-curated benchmark designed to evaluate multimodal large language models' capacities in image aesthetics perception across four distinct dimensions. Evaluations using this benchmark reveal that current MLLMs possess only rudimentary aesthetic understanding, demonstrating a substantial gap compared to human experts, particularly when interpreting AI-generated images and formulating precise aesthetic explanations.
Key insights
- Current MLLMs exhibit only rudimentary aesthetic perception abilities, consistently falling far short of human expert judgment across all evaluated dimensions.
- MLLMs demonstrate varying performance based on image source, generally performing better on natural and artistic images than on complex AI-generated content.
- A critical limitation of MLLMs is their 'unsatisfactory Precision' in aesthetic interpretations, frequently hallucinating details or providing inaccurate explanations, even when overall relevance is present.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.