Mike Tika paints portraits of non-existent people. However, he does not use a brush for this, but the "imagination" of a neural network.
I'm interested in people's faces, you can read a lot in them. I am fascinated by this project, because I love to reflect on who these people would be if they really existed.
Mike spent about nine months developing the Portraits of Imaginary People project, which followed Inceptionism and Groovik's Cube.
GAN, generative adversarial networks
In his project, Mike used generative adversarial networks (GAN):
I started experimenting with GAN in an installation I did with Refik Anadol, where we used this technique to generate imaginary historical documents from a large archive. After finishing the project, I took a fresh look at portraits using this method.
Promotional video:
Let's say you want the GAN to portray a cat. To get started, you need a large number of cat photos. After that, you must prepare a model to create an image of a cat, which would take into account all the features of the animal: whiskers, paws, tail. Just as a baseline dataset for machine learning, Mike used about 20,000 high quality images from Flickr.
But this is only the first step. To get a realistic image of a cat and not a digital sketch, you need to create a second neural network known as a discriminator. While the first neural network (generator) will create pictures of cats, the second (discriminator) will compare the work of the first with real pictures of cats and find out if they are reliable. Based on the results, the system adjusts the generator parameters to make the output image more realistic.
If you use only one network, then the size of the output image will range from 128 × 128 to 256 × 256 pixels. To increase the size of the images, you need to collect several separately prepared GANs so that the next level is a discriminator of the previous one. This step will be able to raise the quality of the image, and the size will vary between 768 × 768 and 1024 × 1024 pixels.
Ultimately, Mike wants to generate 4K photos, but he is currently having a hard time finding a dataset to train the system:
GAN is difficult to train and difficult to control. You need to carefully monitor the input data, make sure that all images are high resolution, have no artifacts and are not drawn. It is difficult to compare different runs with different parameters because there is no good, stable indicator of how well a particular network is performing. And the output image takes a very long time to build. But my project was not created for the sake of accurate results or indicators, but primarily for the sake of art, which should inspire and make you think.
Dmitry Alexandrov