In this post, we compare portraits generated by a neural style transfer algorithm with VGG-19 CNN network and StyleGAN2 trained on FFHQ (Flickr-Faces-HQ). The input for the first algorithm is a minimal sketch, while the input for the second one is the output portrait of the previous system.
| Minimal sketch | → | Neural style transfer with VGG-19 CNN | → | Generative adversarial network StyleGAN2 |
|---|
Cybele. Mountain mother. Nature Deities. 2020
StyleGAN2 is an advanced generative adversarial network trained on FFHQ (Flickr-Faces-HQ), a high-resolution dataset that offers high quality and variation in terms of age, ethnicity and image background. Instead of faces, this project searches in the portrait manifold finding generated images, using photos of nature as an input uncovering its hidden spirits. From European to African, American and Asian mythology, all cultures encompass nature deities embodying natural forces; AI unravels their mystery unifying them in a universal AI mythology.
This project was presented in the AI Art Gallery of NeurIPS 2020, the premier conference in machine learning.
Progressions. Nature Deities. 2020
You can watch the full project including all the nature deities in the video below.
“In the name of the Father.” Byzantine iconography generated by DCGAN.
As Byzantine art depicted the transcendent time and place and it was deeply intertwined with Christian Orthodox faith, it was the exact opposite of classic naturalistic aesthetics. As Ernst Kitzinger, a historian of Byzantine art, stated there was a "dialectic" between "abstract" and "Hellenistic" tendencies in late antiquity. Symbolism was more important than an accurate representation or correct perspective. The figures tended to "hover" in space, occupying symbolically the transcendent rather than the three-dimensional space.
A deep convolutional GAN (DCGAN) was trained on a dataset of Byzantine iconography, which contained hundreds of icons of Christ, Madonna, and various saints. The AI system generates then icons of religious content, evolving the Byzantine style of abstract depiction, where the goal is to represent some "quality" of the meaningful content rather than naturalistic forms. The result is mystic forms of transcendental aethereal attributes, or perhaps an AI interpretation of Byzantine religious figures.
“Non-convergence. Tribute to Jackson Polock.” Generated by DCGAN.
“Animal farm” Generated by DCGAN, 2019.
