Training a neural network in the style of Meybis Ruiz Cruz

The project describes the process of training the Stable Diffusion model to generate portraits in the style of artist Meybis Ruiz Cruz, utilizing Kaggle and a custom dataset. The primary goal is to create a neural network capable of reproducing the artist’s unique style in the generated images.

To train the neural network, a dataset of 50 artworks by the artist was compiled. The artist specializes primarily in portraiture, so the neural network most often generates bust or head portraits.

The artist’s work shows slight stylistic variation: ranging from highly cartoonish to painterly realism. However, all pieces feature «confident» strokes from textured brushes, a certain «angularity, ” and an alternation between flat color fills and gradients.

First, the Marilyn Monroe portrait generation variant was tested. The results showed an inaccurate resemblance to the actress, though the drawing style was considered successful.

Like the author’s work, the generated images vary slightly in proportion. Nevertheless, each drawing effectively mimics painterly brushstrokes—this is noticeable in both the faces and the hair.

After initial tests generating a stylized portrait of an actress, it was decided to incorporate random physical characteristics into the prompts, such as: a man with short dark hair and gray eyes, a sad expression.

The two generated images above were, in my opinion, the most successful options in terms of style fidelity, even despite the slightly «broken» anatomy.