Neural network generations in the style of David Hockney’s work

David Hockney is an artist who, for sixty years, has painted the same joyous subjects: California swimming pools, sunlight on the water, friends in living rooms, and the hills of Yorkshire. His style is immediately recognizable—flat, bright color fields, blue pool spectrums, stubborn wavy lines, deliberately «incorrect» perspective, and signboard-like shadows. Hockney consciously breaks optical accuracy for pictorial pleasure: he paints not how the world looks, but how it feels on a sunny day.

A generative neural network, conversely, learns the world as a dataset: it doesn’t «love» the sun—it averages millions of images. The project’s idea is to confront these two perspectives: training a neural network directly on Hockney’s paintings to see if the statistical machine can absorb not the subject matter, but the joy itself—the sun-drenched color, the flatness, the lightness.

Series Concept: «Sun in a Square»: Each frame is a square «cutout» from the world of Khokhin (the artist often works in series and with square canvases), in which the model collects his techniques into scenes he never painted.

Forty works were selected for the training: swimming pools, portraits, interiors, and landscapes. The variety of subjects helps focus the instruction on the artist’s visual language.

Images are cropped to 1024×1024. Low-quality reproductions, photo collages, and shots containing extraneous objects have been excluded.

The SDXL model was trained using the LoRA method: the base weights were frozen, and additional matrices adapted it to the specific style. The image captions included the trigger «hockney style.»

Training was performed in Google Colab on a T4 using the script diffusers train_text_to_image_lora_sdxl.py.

Parameters: resolution 640×640, 500 steps, batch size 1 with a gradient accumulation of 2, rank 32, learning rate 1e-4, cosine scheduler, 100 warmup steps, fp16, and gradient checkpointing.

The model presented a vibrant palette, flat color blocks, and decorative outlines. The images vary in subject matter, color combinations, and proximity to the original works. The series demonstrates how Hokni’s techniques can be translated into new compositions. Moreover, the expressiveness of the lines and the spatial arrangement require individual evaluation in each frame.

Stable Diffusion XL 1.0 — fine-tuned using LoRA on 40 works by Hockney and used to generate the entire series.

ChatGPT, [GPT6-Astra] — utilized for code debugging and prompt preparation.