Training a generative neural network in the style of Marc Chagall’s work

While working on this project, I trained a generative neural network to create images in the style of artist Marc Chagall. I chose Chagall’s work because his art left a particularly strong impression on me. His paintings are imbued with emotion—they feel light, as if floating in the air. What is depicted in them doesn’t always adhere to logic, which only heightens the interest.

I was curious to see how the neural network would interpret the style and convey the essential elements of Chagall’s works. I used various subjects to ensure the network could handle any theme while maintaining the artist’s style.

For working with the neural network, paintings by Chagall were collected and cropped to a square format. SDXL_DreamBooth_LoRA was used to train the neural network. In the Colab notebook, the GPU is configured, necessary libraries are installed, and the required parameters are set. After all data is prepared, the LoRA is launched, which is necessary for generating new images in the Chagall style («in CHAGALL style»).

During the process, I generated a large number of images which I saved to the drive. Below is a series of images that I believe capture some of the most successful moments, vividly reflecting the core characteristics of Marc Chagall’s work. His paintings are filled with fantasy, where anything is possible—cows fly, and lovers drift above the city.

Analyzing the images generated by the neural network, it is evident that they exhibit the style of Marc Chagall’s work. They do not always copy him directly, but the mood present in them reflects the atmosphere of his art. These images possess a lightness and a sense of flight. Thus, animals and objects often float in the air, and houses are vibrant elements in many of the paintings. The colors in the images, much like in the artist’s work, are bright. Furthermore, the dreamlike or fairytale atmosphere is preserved—everything feels unreal, existing outside the laws of gravity and logic. However, there are issues and errors in human anatomy, particularly with the hands.