Original size 743x1024

Learning Generative Neural Networks in Filonov Style

PROTECT STATUS: not protected
Longread translated automatically

If asked who could have invented neural networks before humanity as a whole considered the potential of their existence, I would say Pavel Filonov. He was the most enigmatic figure of the Russian avant-garde and the founder of analytical art. His method lay beyond the scope of human perception. The canvases truly remind me of a compilation of neural connections, given a new synthesis of previously unseen imagery.

*The works of Pavel Nikolaevich Filonov entered the public domain in 2015.

I found it interesting to train a neural network on his works. I expected that by combining the machine’s thinking with Filonov’s, we would achieve amazing images filled with meaningfulness in its abstraction.

I have gathered all the square paintings by Pavel Filonov that I found into the archive. You can view them here.

0
1/3

Install necessary libraries

big
Original size 2028x266

Initially, it was necessary to install all the required libraries. Additionally, due to a version mismatch in Perf, an update to version 17 and a complete environment restart were required.

Preparation for Study

big
Original size 1138x681
Original size 1318x172

Since the code was generated via Kaggle, I used the built-in function to create a dataset where I uploaded all the photographs. Then, I added them to the working folder cher.

Original size 2034x550

Sign captions were then created for each image. I made sure they are read correctly.

Original size 2036x914

Subsequently, prefixes were generated for each image using a specific function, caption_images. Thus, a prompt was created for each one.

Learning and Model Preservation

Original size 1669x414
Original size 1282x840

On Hugging Face, I obtained a token so that the subsequently trained model could be uploaded to the platform, and I began the training. It took about an hour.

After saving the trained model on Hugging Face, I was finally able to start generating images.

Generation

In search of a function for the trained model, I found it interesting to use it to generate a visual concept for fashion or jewelry photography centered around Russian identity for a potential brand.

Initially, I wrote a fairly brief prompt and received a rather abstract image.

Original size 1024x1024

Kaggle:

photo collage in FILONOV style, fashion photoshoot of a bag against the backdrop of Red Square in Moscow

That’s why I decided to write more specific prompts, and the images became much more detailed.

Original size 1024x1024

Kaggle:

photo collage in FILONOV style, fashion photoshoot featuring a world-famous model wearing a kokoshnik against the backdrop of the Hermitage Museum in St. Petersburg. The photo was taken for a brand presentation in a fashion magazine

Learning Generative Neural Networks in Filonov Style
Project created at 09.11.2025
We use cookies to improve the operation of the HSE website and to enhance its usability. More detailed information on the use of cookies can be fou...
Show more