```html Multimodal Large Language Models and the Evaluation of Aesthetics

Artificial intelligence (AI) is permeating more and more areas of our lives, and art is no exception. Multimodal Large Language Models (MLLMs), which can process both text and images, open up new possibilities for interacting with and evaluating artwork. A recent study by the Hong Kong Polytechnic University investigates how MLLMs assess the aesthetic quality of artwork and to what extent this assessment aligns with human preferences.

The Challenge of Aesthetic Evaluation

The evaluation of aesthetics is inherently subjective and complex. What is perceived as beautiful or appealing depends on individual preferences, cultural background, and many other factors. Previous attempts to capture aesthetics with AI often focused on purely visual features and neglected the cultural context and emotional impact of a work of art. This led to results that often did not match human assessments.

MM-StyleBench: A New Benchmark for Artistic Stylization

To investigate the capabilities of MLLMs in aesthetic evaluation, researchers at the Hong Kong Polytechnic University developed the MM-StyleBench dataset. This dataset contains a variety of images and text descriptions with detailed attribute annotations, enabling the evaluation of artwork stylization. MM-StyleBench provides a basis for comparing different MLLMs and their ability to recognize aesthetic preferences.

ArtCoT: A Method for Improving Aesthetic Evaluation

The researchers also developed a new method called ArtCoT, which improves the ability of MLLMs to perform aesthetic evaluations. ArtCoT is based on task decomposition specifically tailored for art evaluation. The method comprises three phases:

1. Content Preservation Evaluation: The MLLMs assess the extent to which the content of the original artwork is preserved in stylized versions.

2. Style Fidelity Evaluation: The MLLMs evaluate how well the stylized images correspond to the given artistic style.

3. Art Critique: The MLLMs provide a detailed critique that connects visual features with art-specific knowledge.

Through this structured approach and the use of concrete language, MLLMs can improve their ability to perform aesthetic evaluations and reduce so-called hallucinations, i.e., the generation of inaccurate or misleading information.

Results and Outlook

The results of the study show that ArtCoT significantly improves the agreement between the evaluations of MLLMs and human preferences. The method enables the MLLMs to generate more nuanced and meaningful evaluations that better reflect the complex nature of aesthetic judgments. These research findings offer valuable insights into the application of MLLMs in the art field and can be used for various applications such as style transfer and the generation of artistic images.

The development of MLLMs that can recognize and evaluate aesthetic qualities opens up new possibilities for the art world. From supporting artists in developing new styles to personalized recommendations of artwork – the combination of AI and art promises exciting developments in the future. The research from the Hong Kong Polytechnic University makes an important contribution to this field and demonstrates the potential of MLLMs for the evaluation of aesthetics.

Bibliography: - https://www.chatpaper.com/chatpaper/fr/paper/100585 - https://arxiv.org/abs/2308.04152 - https://www.chatpaper.com/chatpaper/fr?id=4&date=1736956800&page=1 - https://medium.com/to-data-beyond/important-computer-vision-papers-for-the-week-from-30-12-to-05-01-6ae0d433c0bb - https://github.com/friedrichor/Awesome-Multimodal-Papers - https://2024.emnlp.org/program/accepted_findings/ - https://arxiv.org/abs/2205.11916 - https://aclanthology.org/2024.naacl-long.117v2.pdf - https://2024.aclweb.org/program/finding_papers/ - https://www.researchgate.net/publication/360834082_Large_Language_Models_are_Zero-Shot_Reasoners ```