Towards AI That Truly Understands Art: Introducing the ArtiMuse Aesthetic Understanding Model

Introduction

ArtiMuse is an image aesthetic model that outputs both a structured textual critique and a composite aesthetic score. The code, model weights, and the ArtiMuse‑10K dataset are fully open‑source.

Motivation

Existing aesthetic evaluators either provide only a numeric score without explanation or generate generic, non‑quantitative descriptions, and they are trained primarily on Western corpora, limiting performance on Eastern art forms such as Chinese ink painting and calligraphy.

Key Contributions

ArtiMuse‑10K benchmark

ArtiMuse‑10K contains 10,000 high‑quality images across five major categories—graphic design, 3D design, AIGC‑generated images, photography, and traditional Chinese art—and 15 sub‑categories. Professional annotators with 3–30 years of experience labeled each image using a “multi‑dimensional expert commentary + composite aesthetic score” scheme covering eight aesthetic dimensions (composition, visual elements, originality, creativity, etc.).

Unified quantitative‑and‑qualitative framework

The model generates a detailed analysis for each of the eight dimensions and aggregates them into an overall aesthetic score, enabling both interpretability and precise ranking.

Token‑As‑Score strategy

Instead of decomposing continuous numbers into separate tokens, the method maps continuous scores onto existing tokens in the language model’s vocabulary, dramatically reducing quantization loss and improving stability of numeric predictions.

Evaluation

Benchmarked against GPT‑4‑Vision, Gemini, Q‑Align, and UNIAA on standard aesthetic rating datasets such as AVA and PARA, ArtiMuse achieved higher accuracy and robustness. On the newly introduced ArtiMuse‑10K benchmark it outperformed all competitors by a large margin.

Qualitative examples show the model’s ability to detect subtle technical details in high‑quality artworks and to pinpoint flaws—e.g., unnatural texture handling—in low‑aesthetic images.

In tests on Chinese traditional paintings, the model correctly identified brushstroke tension, ink shading, and the symbolic meaning of “松鹤延年” (pine and crane symbolizing longevity), demonstrating reduced Western‑centric bias.

Resources

Online demo: https://artimuse.intern-ai.org.cn/

GitHub repository: https://github.com/thunderbolt215/ArtiMuse

Project homepage: https://thunderbolt215.github.io/ArtiMuse-project/

ArXiv paper: https://arxiv.org/abs/2507.14533