How A.I is reinventing the artistic process in NFTs.
Introduction
NFTs (Non Fungible Tokens) are immutable pieces of data stored in the blockchain. Recently they have been an especially popular way to store art. Images, GIFs and videos are now saved forever in the blockchain as NFTs and can be easily traded from hands to hands almost instantaneously.
Storing art as NFTs offers several advantage over traditional mediums. The first one mentionned earlier is speed and ease of transaction. There is no need for physical transactions and storing a physical art piece. The fact that all transactions happen over the blockchain means that anyone around the world can potentially buy any artwork produced, anywhere in the world. The second advantage of NFTs is easy authentification. With NFTs anyone is able to quickly verify the authenticity of any piece on the blockchain.
Thanks to all those innovations, it has never been easier to connect artists and collectionners before and that’s why the NFT art market will without a doubt take a larger and larger share of the global art market. The demand for NFT art is growing day by day but one challenge for the artist is that of “scaling”, to provide unique, inspiring and beautiful art pieces for all this growing interest. This is where artificial intelligence can help. One such technique is GANs (Generative Adversarial Networks).
GANs in a nutshell
GANs are a special architecture of neural network that belongs to the family of generative models. This means that given a training dataset, they are able to produce output similar to the training input. Using a large dataset of paintings (Wikiart is the most comprehensive source), one is able to train GANs to reproduce the paintings seen as input.
How does it all work?
We will give here a simplified overview of the principles of training to make it understandable with very little to no knowledge of machine learning. GANs are actually made of two neural networks. One is called the generator model, this is the one that will learn to “paint”. The other one is called the discriminator model. Those two models are trained in sync. The discriminator model receives as input some paintings from the dataset and some paintings artificially generated by the generator model. The discriminator must determine whether it is looking at a true painting or a fake painting artificially generated. Conjointly the generator must learn how to fool the discriminator and make it think that the paintings produced are real. At the end of the training process the generator learns how to pass the discriminator’s test, i.e its paintings are now indistinguishable from the true paintings from the dataset.
Applications
GANs can learn how to generate any kind of data, it all comes up to the dataset used as input and the architecture chosen for the neural networks. Hence, the applications are very diverse.
For example, as demonstrated in the paper “Progressive Growing of GANs for Improved Quality, Stability, and Variation”, GANs can be used to generate perfect photographs of human faces. You can have a look at this-person-does-not-exist.com. The results are very unsettling.
And there is of course the application of generating scalable unique art that we mentionned throughout the article. GANs trained properly are able to generate thousands of unique art pieces from a very wide variety of styles. This is precisely the technique that we use at Canvax, to generate beautiful art stored as NFT on the Solana Blockchain. Using datasets from Wikiart, from Cubism, Impressionism, Romanticism and Ukiyo-e, we have trained art to generate classic paintings with a wide variety in the art produced.
