Dmitri Cherniak

2020–2023

gen2GAN

With Helena Sarin

Begun in 2020, gen2GAN brings together Dmitri Cherniak’s generative algorithms and Helena Sarin’s work with generative adversarial networks. Cherniak’s software produced thousands of uncurated images, which became the training material for neural networks constructed and tuned by Sarin.

The artists selected the final works together. The collaboration moves from an artist-built system to an artist-built dataset and model, exploring how one form of automation can transform another.

Began
2020
Artists
Dmitri Cherniak and Helena Sarin
Process
Generative algorithms and GANs
Black and white striped forms with vivid red interiors on a pale ground.
Helena Sarin and Dmitri Cherniak, Red October, 2020.
Red October at Sotheby’s

From the series

Selected works

Orange and blue concentric circles, rippling black lines, and delicate threads across a white field.
Helena Sarin and Dmitri Cherniak, K-Meanearest Neighbors, 2020.
K-Meanearest Neighbors at ENG.ART
Irregular black and red forms arranged across a pale field.
Helena Sarin and Dmitri Cherniak, A Big Band, and All That Jazz, 2020.
A Big Band, and All That Jazz at Sotheby’s

The process

From generative code to GAN

Cherniak’s JavaScript generates SVG compositions, which are rendered as PNGs without curation. Sarin uses these images to train GAN models. The models and their outputs are then curated to form the collaboration’s final works.

Five stages of the collaboration: JavaScript code; SVGs generated in under 100 milliseconds; rendering to PNG without curation; GAN model training; curated models and output.
gen2GAN collaboration diagram. From generative code to curated GAN output.

The dataset

Many inputs, one model

Each GAN was trained on thousands of Cherniak’s uncurated outputs. The 25 compositions shown here come from one dataset of 2,000+ generated SVGs, alongside an output from the model trained on it.

Generated training image 1: black and red bars with circles on gray.Generated training image 2: black and red bars with circles on gray.Generated training image 3: black and red bars with circles on gray.Generated training image 4: black and red bars with circles on gray.Generated training image 5: black and red bars with circles on gray.Generated training image 6: black and red bars with circles on gray.Generated training image 7: black and red bars with circles on gray.Generated training image 8: black and red bars with circles on gray.Generated training image 9: black and red bars with circles on gray.Generated training image 10: black and red bars with circles on gray.Generated training image 11: black and red bars with circles on gray.Generated training image 12: black and red bars with circles on gray.Generated training image 13: black and red bars with circles on gray.Generated training image 14: black and red bars with circles on gray.Generated training image 15: black and red bars with circles on gray.Generated training image 16: black and red bars with circles on gray.Generated training image 17: black and red bars with circles on gray.Generated training image 18: black and red bars with circles on gray.Generated training image 19: black and red bars with circles on gray.Generated training image 20: black and red bars with circles on gray.Generated training image 21: black and red bars with circles on gray.Generated training image 22: black and red bars with circles on gray.Generated training image 23: black and red bars with circles on gray.Generated training image 24: black and red bars with circles on gray.Generated training image 25: black and red bars with circles on gray.
25 of the 2,000+ generated training images.
A GAN output of black and red vertical bars and strokes scattered across a pale gray field.
Red Desert, After CalderOutput from the model trained on this dataset

Try it

Train a tiny GAN

A GAN is two neural networks playing a game. The generator turns random noise into an image. The critic, or discriminator, is shown real training images and the generator’s fakes, and gives each one a probability that it is real. After every round, the critic is adjusted to tell them apart better, and the generator is adjusted toward whatever the critic scored as more real.

This one is tiny. It trains in your browser on the 2,000+ images from the dataset above, each reduced to a grid of 16 × 16 cells in its four colors. Sarin’s models were far larger and worked at full resolution; this is a sketch of the same game.

This demo trains in your browser and needs JavaScript.

The models

Five models

The collaboration produced far more than the works that were released. Each of five models was trained on a different generative system, and each developed its own visual language. One output from each is shown here.

GAN output “The Cat and the Mouse” from the A Slight Lack of Symmetry model. Looping, banded forms dissolving into soft organic shapes.
The Cat and the MouseA Slight Lack of Symmetry model
GAN output “Escalier” from the Autoneoplasticism model. Red and blue blocks and a stair-like diagonal of red lines over yellow and white stripes.
EscalierAutoneoplasticism model
GAN output “Love Your Experiments” from the Good Russian Jew model. Black and red bars smeared into horizontal strokes.
Love Your ExperimentsGood Russian Jew model
GAN output “Remod 02” from the Stained Glass model. Colored panels and outlines dissolving into green and yellow fields.
Remod 02Stained Glass model
GAN output “Ruby Sunday” from the A Slight Lack of Symmetry II model. Red and black blob-like forms with banded outlines scattered across a white field.
Ruby SundayA Slight Lack of Symmetry II model

Publication

gen2GAN, the book

The collaboration also took the form of an artist’s book, published in 2023 in an edition of 200. Its spreads place Cherniak’s generated images alongside Sarin’s GAN outputs, making the exchange between the two systems visible on paper.

Published by DLC Arts LLC, with a foreword by Casey Reas and an introduction by A.V. Marraccini.

Book credits

Artists
Helena Sarin and Dmitri Cherniak
Curated by
Sofia Garcia
Foreword
Casey Reas
Introduction
A.V. Marraccini
Publishing director
Eli Rosenbloom
Art director
Emwhi Nguyen
Designer
Yen Ho
Copy editor
Eden Rosenbloom
Production
Functional Brands LLC
Editorial assistant
Huy Vu
Layout assistant
Kristi Huynh
Printing
Conveyor Studio
Coverslip
Shapco Printing Inc
The gen2GAN book cover, showing a grid of generative compositions transitioning into GAN outputs, photographed on a pale wooden surface.
gen2GAN, 2023. Published by DLC Arts LLC.
Published
2023
Edition
First edition · 200 books
Publisher
DLC Arts LLC
ISBN
979-8-218-08636-7
Printed and bound
USA