Artistic Style Imitation with Generative Artificial Intelligence

Artificial intelligence has spread very fast in recent years, leveraging various kinds of architectures. Among the several applications, style transfer is one of the most promising. The purpose of this study is to evaluate the state-of-the-art models in the capability to generate paintings of artistic movements from 1500s to the first half of 1900s. To achieve this, we created a labeled dataset of 953 images using 12 state-of-the-art generative models, including DALL-E 3, Midjourney, Stable Diffusion, and Leonardo Phoenix. In addition, we designed two surveys to assess both the authenticity of the generated images - if they could be misclassified as human-made - and their adherence to the prompt. As a result, some models prioritize aesthetic quality over strict adherence to the prompts, while others sacrifice authenticity for greater accuracy. Ideogram 2.0 showed strong performance both in authenticity and prompt adherence combined while Auto-Aesthetics is the one that performs the worst. Despite the rapid advancements in generative models, our results show areas for improvement. Generated artworks were classified as human-made less than 30% of the cases, and the models demonstrated better performance in replicating artworks of recent artistic movements than older ones. This work contribute with an analysis of the capabilities and the limitations of state-of-the-art text-to-image generative models and a dataset of synthetic images which can be used for further research.