GANs Exploration:
Najeon Chilgi
Najeon Chilgi is Korea’s traditional craft of Mother of Pearl Lacquerware. This work is an exploration on various image generation methods. These methods include:
1) DCGAN + ISR with Google Colab
2) Style GAN with RunwayML
3) StyleGAN with Google Colab
machine learning ✔experimental art ✔
About This Work
It’s an exploration on GAN workflows and models to determine the pros and cons of different approaches.
-
Google Colab, Photoshop, Python, Runway ML
-
An image dataset was scraped and preprocess to amass a collection of 3,000+ images.
-
Deep Convolutional Generative Adversial Network (DCGAN) is limited in that the input dataset must contain 128x128 pixel images, severely depressing the image quality.
DCGAN’s synthesized images were appealing for their more saturated colors. In an attempt to improve the image quality, the synthesized images were ran through Image Super-Resolution (ISR) model.
-
RunwayML is a platform for artists to use machine learning tools without any coding experience.
To create synthetic images using the platform, a StyleGAN and the pretrained ‘Flickr Faces HQ’ model were used for the najeon chilgi dataset to transfer learn against. The model trained for 2000 steps.
-
Colab is a web based IDE for python with free GPU recourses.
Between RunwayML and Colab, Colab requires more involvement in the overall process and allows for more cusomization. The Style GAN2 model also trained the the Najeon Chilgi model by transfer learning against a dataset of faces.