This Glaze Does Not Exist.
Curated images with 𝜓 1.0 generated using the network trained for 782 iterations against the Glazy glaze test dataset.
In December 2020, the Nvidia team released version 2 of StyleGAN which offered better performance than version 1, and released a companion paper, "Analyzing and Improving the Image Quality of StyleGAN".
Excellent article about running StyleGAN2 by Zalando Dublin here: StyleGAN v2: notes on training and latent space exploration
Out of curiosity and as a test run of StyleGAN2, I created a dataset of about 15,000 public images from Glazy I manually removed a number of images that included hands & fingers, complex background elements, and other disqualifying characteristics. However, the remaining images still included a wide variation of not only colors and surfaces, but also shapes of glaze test tiles. I used ImageMagick convert to precisely resize the images to 512x512 pixels in dimension, then used identify to verify that all images were in the correct sRGB colorspace. Finally I ran StyleGAN2's dataset_tool.py to create the multi-resolution datasets.
I again used Google Cloud's AI Platform on a server with a single Nvidia V100 GPU. I encountered some issues with the server environment, my advice is to just follow the Requirements section of the StyleGAN2 documentation and ensure you are running the exact same versions of CUDA, CuDNN, TensorFlow, Python, etc. Otherwise you might spend a lot of time fixing dependencies and dealing with version conflicts like I did.
I trained a new network from scratch using configuration "E":
python stylegan2-master/run_training.py --num-gpus=1 --data-dir=datasets --config=config-e --dataset=glaze --total-kimg=10000
Given this poor dataset, I was not optimistic. But after only a few hundred iterations the results were already very promising. At around 800 kimg the images were already good enough as a proof of concept, and I stopped training.
