Kustodiev’s Visual Code

1. «A painting in the style of Kustodiev: a beautiful, voluptuous Russian merchant’s wife in a bright red sarafan standing by a window with blooming lilacs.»2. «A painting in the style of Kustodiev: a beautiful Russian girl wearing a traditional kokoshnik and holding a bouquet of flowers.»3. «A painting in the style of Kustodiev: a portrait of a wealthy Russian merchant with a large beard and kind eyes.»4. «A painting in the style of Kustodiev: a portrait of a voluptuous Russian beauty with rosy cheeks wearing a traditional folk costume.»5. «A painting in the style of Kustodiev: a beautiful, voluptuous woman with rosy cheeks wearing a bright exotic outfit made of tropical flowers and leaves, gracefully standing among lush jungle vegetation, palm trees, and colorful parrots.»6. «A painting in the style of Kustodiev: a colorful spring fair with carousels, balloons, and a cheerful crowd.»7. «A painting in the style of Kustodiev: a romantic winter evening, a couple riding in a horse-drawn troika under a starry sky.»8. «A painting in the style of Kustodiev: a majestic sailing ship with billowing white sails sailing across a calm turquoise sea at sunset, surrounded by flying seagulls and distant islands.»9. «A painting in the style of Kustodiev: a quiet Russian riverside at sunset, blooming willow trees, wooden boats, and a distant monastery reflected in the calm water.»10. «A painting in the style of Kustodiev: an endless golden wheat field beneath a bright blue sky with fluffy white clouds, with a church featuring golden domes and a small village in the distance.»

For training the neural network, artworks by the Russian artist Boris Kudostyov were selected. The generated series features various scenes related to Russian culture and everyday life: portraits of people in traditional costumes, fairs, festive celebrations, rural landscapes, architecture, and scenes of urban life. Despite the diversity of subjects, all images share a visual language characteristic of the artist’s work.

The primary objective was not to reproduce specific Kudostyov paintings, but rather to capture the nuances of their artistic expression and translate those qualities onto new subject matter. The works retain a rich and contrasting color palette, high degree of ornamentation, abundant detail, and a sense of festivity. Special attention is paid to red, blue, yellow, and green hues, which create a vibrant and cohesive visual image.

A defining feature of the resulting series is its dense composition. Space is filled with people, objects, architecture, vegetation, and decorative elements. The foreground and background actively interact, lending the images depth and a sense of fullness. Even with a high volume of objects, none appear accidental—they all contribute to establishing the overall atmosphere of the scene.

The human figure also plays a significant role. The works are characterized by expressive figures, voluminous silhouettes, rosy faces, traditional attire, and pronounced decorative elements in the costumes. When generating new characters, these features were used as one way to maintain a connection to the visual material on which the model was trained.

Variability within the series is achieved by changing the plot, the environment, and the season.

From a generation perspective, the model adequately renders the overall color palette, decorative quality, compositional density, and festive atmosphere characteristic of the training material. However, individual fine details, complex interactions involving numerous characters, and certain architectural elements are reproduced less consistently. Therefore, the final series was hand-selected: from several options, the images most accurately corresponding to the specified visual direction were chosen.

The primary result was achieved by training the model on the selected dataset and subsequently working with text prompts. While the subject matter changed when generating new images, the key visual characteristics of the training material were maintained.Consequently, a series of new images were created that do not copy individual works by Kustodiev, but rather utilize the characteristic principles of color, composition, characters, and decorative details found in his work. Thus, the experiment demonstrates how a trained model can transfer the visual characteristics of a specific artistic material onto new narratives, preserving the recognizable style while altering the content of the image.

Learning Process

The Stable Diffusion XL model was used for this project, incorporating DreamBooth and LoRA within the Google Colab environment.

The process involved several stages: dataset preparation, model training, LoRA integration, and subsequent image generation.

Stable Diffusion XL — used for image generation.
ChatGPT — used for text and prompt work.