Original size 1140x1600

Training a generative model in the style of Redon

PROTECT STATUS: not protected
Longread translated automatically

Navigation

— Project Idea — DadaSet — Learning Process — Notebook Link — Final Image Series — GenAI — Conclusion

Project Idea

For my project, I chose the works of the French Symbolist painter Odilon Redon (1840–1916), focusing specifically on his color period. During this time, the artist worked with pastels, oil paints, and color lithography, gradually moving away from dark, monochromatic graphic works. His later pieces are characterized by a bright yet restrained palette, fantastical themes, and symbolic imagery.

Project Goal — to train a neural network to recognize and reproduce the characteristics of Redon’s colorful artistic language: his signature color combinations, the soft texture of pastel, and his symbolic imagery.

big
Original size 1102x396

Redon’s Works

big
Original size 1714x2380

Compiled Dataset

The dataset includes 35 color works by Odilon Redon. It spans different stages of the artist’s color period: from 1890s pastels to 1910s decorative compositions. The selection criteria considered the diversity of subjects and how expressively the characteristic features of Redon’s palette and composition appear in each piece.

Before uploading, the images were converted to a uniform square format and pre-processed.

Original size 1102x396

Redon’s Works

The Learning Process

The training utilized a dataset consisting of 35 artworks from Odilon Redon’s color period. All images were in a square format, which simplified data preparation and contributed to stable training.

The model was trained on a GPU in Google Colab.

Original size 967x702

To train the model on the style, I used BLIP for automatic image description generation. I added a stylistic prefix with the trigger token ODRNCLR to each description to link all images to a single visual style.

Original size 1402x455

After completing the training, I authorized myself on Hugging Face and uploaded the model weights to my repository to save them and use them in the future.

Original size 1095x179

Training was conducted using the Stable Diffusion XL model.

Original size 792x517

To generate the images, prompts were used with the trigger token ODRNCLR and descriptions reflecting the characteristic traits of Odilon Redon’s style. As a result, a series of images was created, conveying the key features of his visual language.

Original size 971x331

1. «painting in ODRNCLR style, serene female profile surrounded by flowers, symbolist colored pastel, blue and ochre palette, mystical atmosphere» 2. «painting in ODRNCLR style, sleeping female figure among irises and lilies, soft turquoise and ochre palette, misty light»

Redon created images on the edge of reality and dreams—with a soft, pastel texture, a subdued yet expressive palette, and symbolic motifs. The goal was not merely to stylize images, but to convey the very «Redonian vision» to the viewer.

Original size 800x800

«painting in ODRNCLR style, winged spirit hovering above dark flowers, colored pastel, soft violet haze, mysterious light»

The generated works feature the artist’s characteristic color palettes: warm ochre, golden, and terracotta hues coexist with deep blues, purples, and emerald greens. The compositions are most often built on close-up shots—objects are partially cropped by the frame’s edge, while the background remains minimalist.

Training a generative model in the style of Redon
Project created at 26.09.2026
adcr.dafes
The Art Directors Club of Russia sets the benchmark for creative excellence in design and advertising, serving as a central hub where industry professionals connect. Since 2004, the Club has hosted the ADCR Awards—a premier national competition celebrating creativity, design, and advertising. The core mission of the awards is to champion agencies and independent creators making their mark in the local market. By presenting this honor on behalf of the professional community, the Club reaffirms its endur... Show more
adcr advertising
adcr design
adcr students and young creators