The City Chose Me
Egor Sizov
Egor Sizov
The project was made in the ACT.AI workshop of the Hybrid Workshops of the Art&Science Center of ITMO University
Description
I was born in the millennium in already post-Soviet Russia, but the childhood and youth of our generation were mostly spent in yet Soviet urban environment.
With the help of portraits of my friends generated by an adversarial neural network trained on a dataset of images of the Soviet urban environment and architecture, I want to reflect the influence of the Soviet urban environment on the personal and collective portrait of my generation, to show the inextricable connection between a person and a city.
Technology
I used the adversarial neural network StyleGAN2 for images of 256x256 pixels. Generated images was upsampled to 2048x2048 pixels using the Real-ESRGAN adversarial neural network.
Dataset
The dataset is based on images of the Soviet urban environment of the Kuibyshev (now Samara) region. I used photographs and postcards collected by enthusiasts on the public resource PastVu. The original images were cropped to a square, resized to 256x256 pixels, converted to grayscale and increased contrast.
Training
The pretrained model was retrained on three datasets:
Dataset #1 consists of about 100 black and white images. At this stage, the model has trained to reproduce the general outlines of buildings.
Dataset #2 consists of about 200 black and white images with high microcontrast, dominated by the textures and rhythm of pannel building windows.
Dataset #3 consists of about 150 color images. As a result of this stage, the model regained the ability to reproduce colors, which was lost during the previous two stages.
The project presents images generated by the model after the second and third stages of additional training.
Reference images
The reference portraits for implement the neural network were taken from the social networks of the project's heroes.
I was born in the millennium in already post-Soviet Russia, but the childhood and youth of our generation were mostly spent in yet Soviet urban environment.
With the help of portraits of my friends generated by an adversarial neural network trained on a dataset of images of the Soviet urban environment and architecture, I want to reflect the influence of the Soviet urban environment on the personal and collective portrait of my generation, to show the inextricable connection between a person and a city.
Technology
I used the adversarial neural network StyleGAN2 for images of 256x256 pixels. Generated images was upsampled to 2048x2048 pixels using the Real-ESRGAN adversarial neural network.
Dataset
The dataset is based on images of the Soviet urban environment of the Kuibyshev (now Samara) region. I used photographs and postcards collected by enthusiasts on the public resource PastVu. The original images were cropped to a square, resized to 256x256 pixels, converted to grayscale and increased contrast.
Training
The pretrained model was retrained on three datasets:
Dataset #1 consists of about 100 black and white images. At this stage, the model has trained to reproduce the general outlines of buildings.
Dataset #2 consists of about 200 black and white images with high microcontrast, dominated by the textures and rhythm of pannel building windows.
Dataset #3 consists of about 150 color images. As a result of this stage, the model regained the ability to reproduce colors, which was lost during the previous two stages.
The project presents images generated by the model after the second and third stages of additional training.
Reference images
The reference portraits for implement the neural network were taken from the social networks of the project's heroes.
Dataset#1
Dataset #3
Dataset #2
