Immortal Author
The ethical implications of using generative neural networks to mimic and copy an author’s style can be assessed in various ways. However, among other things, this process is permeated by the idea of the creator’s digital immortality.
During the development of this project, I immersed myself in studying the work of the painter John Atkinson Grimshaw. This renowned Victorian-era artist left behind a considerable legacy—over one hundred and fifty paintings. The majority of these consist of meticulously detailed urban and natural landscapes of the English countryside. Great attention was paid to the play of light—on the walls of buildings, wet cobblestones, the dense haze of fog, or the thousands of leaves on autumn trees.
I became curious—what if he could travel through the world and through time? How would modern architectural landmarks look in his paintings?
Source Dataset
I selected 20 pieces by the artist, taking a 1:1 square crop from each. Given that most of the originals are in a vertical aspect ratio, the challenge was also ensuring that the selected fragment did not critically disrupt the author’s composition.
To highlight the author’s stylistic characteristics, a generative model, Stable Diffusion, was fine-tuned on this dataset using the Dream Booth and LoRa methodologies.
Hugging Face was utilized to apply these tools.
Pre-training Results
After pre-training the model, new images of contemporary architectural landmarks and industrial landscapes were generated using specific prompts. The prompts, complete with object examples, were created using ChatGPT.
The resulting images adopted the stylistic characteristics of the artist. The main compositional solutions, the nuances of light and shadow, and the significant attention paid to the influence of atmospheric effects and weather on the viewer’s «optics» are easily discernible. Nevertheless, the generated images became «stuck» on a green color palette, which indicates the need for more meticulous tuning and control during the pre-training phases.
