Paradising is a digital video artwork released on Sedition. Trained with images tagged with the keyword paradise scrapped from public domain archive, this video artwork seeks to explore the latent space of historical imaginaries of paradise through the use of machine learning generative adversarial network algorithms.

View on Sedition

I scraped all the images from Internet Archive that contains “paradise” in the metadata with Internet Archive python API. Then I used data-set tool to treat the dataset by resizing, cropping and border-handling. Combined with manual filtering out the images not in the public domain or with potential copyright issue, I ended up with a set of 1255 images.

Then I used Google Colab StyleGan2-ADA notebook to train all the images over the span of two weeks with 5000kimg. I had to resume the training almost every 20 hours with a yield of 400kimg. It was quite often that the some cycles of the training would finish way before 20 hours.

fakes (stylegan2 model) initial

reals (partial dataset)

model training in progress 1

model training in progress 2

After the model is trained, I imported the model into StyleGAN2-ADA-PyTorch notebook to experiment with different interpolations, truncation, and random seed.

Some making-of exports:

noise loop, diameter 0.9, random seed 100

circular loop, diameter 500, truncation 1.8, random seed 200