Modeling Constrained Creativity: A Computational Framework for Interaction-based Painting Evolution

Sociologists have long recognized that artists are the product of social and spatial constraints. Yet, bridging the constrained process of artistic creativity with macro-level evolution of art genres remains methodologically difficult because of three related problems: measuring continuous stylistic features beyond genre labels, tracing dynamic learning processes from earlier to later works, and scaling micro-level learning behaviors into aggregate trajectories of stylistic change. To address these problems, this paper introduces a computational framework combining statistical modeling and generative agent-based simulation. First, I construct directed painting-pair comparisons and measure stylistic learning through the similarity between embeddings derived from a deep learning model from painting images. Then, I use linear regressions to model the learning rates by ties (self, first-order/direct, and second-order/indirect) and geographical distances between artists based on the online bibliographies. Finally, the empirical parameters derived from these models are then integrated into a generative agent-based simulation, in which artist agents iteratively generate images, learn from other agents, and update their stylistic tendencies. Applying the framework to the 1920s Paris-centered artistic community, using more than 300,000 directed dyadic artwork comparisons, the paper finds that stylistic learning is shaped by differentiated effects of social ties and spatial constraints. By adopting computer vision methods of both representation learning and generative models, the framework advances a scalable method for testing how micro-level social learning can accumulate into broader cultural evolution.