AI is powered by algorithms on computers to learn from patterns to understand what's what. In turn, AI can be made to have some sort of 'imagination'.
The 'doesnotexist' suffix has been used for quite some times to show how AIs can imagine non-existing things and characters.
From 'This Fursona Does Not Exist', to 'This Meme Does Not Exist' and 'This Word Does Not Exist' and many more.
Not to mention the 'This Person Does Not Exist' that was considered one of the earliest that started the trend, and the 'These Nudes Do Not Exist' that took this 'doesnotexist' thing NSFW.
This time, the saga continues, with 'This Pony Does Not Exist'.
This particular AI is similar to other 'doesnotexist' AIs, but spews out an endless stream of new pony characters from My Little Pony.
Adorable AI ponies have arrived!
The model is transfer learned from the furry model, but with 2x the training data, and 2x the network width/height.
Enjoy.https://t.co/IWWAEvEHFg— arfa (@arfafax) July 15, 2020
The thisponydoesnotexist.net site was created by a developer known as Arfa, the same mind behind 'This Fursona Does Not Exist'.
According to the site, the training dataset consisted of around 104,000 SFW images from Derpibooru, considered one of the world’s largest imageboard for My Little Pony artworks. The images were cropped using the open-source project MLP-Face-Dataset, to then aligned to the faces using a custom YOLOv3 network.
Arfa built the system by using the same off-the-shelf machine learning product, TensorFlow and Javascript, and trained his model using a generative adversarial network (GAN).
This GAN pits two neural networks against each other. The first is called the generator, which produces new images. The second is the discriminator, which determines whether the image created by the the generator is considered genuine by comparing it with the original dataset.
Pitted against each other, the generator continued to refine its designs until the discriminator decides that the results can pass as the real thing.
The model used transfer learning to fine tune the final model from Arfa's existing 'This Fursona Does Not Exist' on the pony dataset for an additional 13 days (1 million iterations), using 512x512 resolution. Arfa then fine-tuned the model by scaling the images up to 1024x1024 resolution using model surgery, to then trained it again for an additional 200,000 iterations to produce the final 1024x1024 model.
The result is a model capable of showing endless ponies that don't exist.
Arfa showcases the model on the website, where the model spits out non-existing ponies in an endless side-scrolling streams of images inside grids.
While some of the ponies resemble the images they were trained on, Arfa said that none of them are exact copies:
The images the model creates are non-copyrightable, said Arfa
"I claim no legal ownership or rights to any of the images generated by this AI. Please use them responsibly."