The (Real) State of the Art — How Artificial Intelligence Is Redefining NFTs and the Creation of Art.
The world of art has been recently reshuffled by technology. The emergence of NFTs (Non-Fungible Tokens) has allowed artist to expand the reach for their works and sell them in fully electronic format. One great advantage of NFTs is that the ownership of the artwork is recorded on the blockchain — usually ethereum — and cannot be tempered with. This gives security to the owners that the artwork belong entirely to them. Art proposed as NFT is usually crafted with creative software like Photoshop or Blender.
Another revolution in the world of art is the possibility to use artificial intelligence to create art. Since the advent of deep learning and more precisely GANs (Generative Adversarial Network) models from 2014, it has been possible to train deep neural networks on art images and have them create original art.
In this article we explain how AIart works and how it is now merged with blockchain technology to create the next generation of NFTs.
1- What is A.I. art and how has it evolved so far?
A.I. is a catch all name which groups different techniques used in computer science. Most of those techniques are based on deep neural networks. Neural networks are simply algorithms which function in two time steps. In the first time step the algorithm (a.k.a the computer code) is given examples of images with labels (for examples images of cat and dogs) and it is trained to minimize an error function so that the code can distinguish between cats and dogs. In a second step, given new, never seen before pictures, the algorithm is able to recognize which ones are cats and which ones are dogs. Usually the precision is not perfect, but close to 95 percent. There are technical names for the models used to achieve those results such as CNNs (convolution Neural Networks) or ViT (Visual Transformers) and we refer the reader to specific articles about those for more details.
A natural question after understanding this is as follows: Since the algorithm can distinguish between different categories of objects, can it actually create new original images?
It turns out this is possible by using GANs (Generative Adversarial Networks). This deep neural networks architecture had been introduced by Ian Goodfellow and al. Back in 2014. The idea is using two deep neural networks competing against each other in successive turns. One network (the generator) is trying to fool the other one (the discriminator) and constantly improves the quality of the image it creates based on the feedback provided by the discriminator.
Naturally the images created need to refer to some “base” images which the generator is trying to imitate. This means that each GAN training run needs to be associated with some dataset. If a GAN is trained on images of cats, it will be able to create images of cats (but not portraits), so there is a direct relationship between the dataset and the nature of the output of the GAN.
Initially GANs were trained on the infamous MNIST dataset of digits, and back in 2014 those were able to create new original drawings of new digits.
The next step in this evolution was to train the GANs on some art datasets and see what the results would look like. Back in 2017–2018 some artists such as Obvious Art, D’AgostinoAI or Robby Barrat started experimenting with GANs with more or less aesthetic success.
Critics of Robbie Barrat wrote (Bonnie Burton in CNET): “The results are surreal. Barrat posted many of the final pieces of artwork — which can only be described as surreal, blobby, swirly naked women — on Twitter. It’s almost like a very intoxicated Salvador Dali and a dizzy Picasso joined forces to make art. …Barrat’s AI-assisted artwork isn’t exactly sensual. In fact, most of the nudes look like they are melting on a very hot day.”
Initially critics were harsh and indeed some works just didn’t meet the aesthetic criteria for solid art. it was experimental and frankly sometimes embarrassing. On the other hand, other artists started putting out quality works, by using different techniques and more computing power.
In 2020 new form of GANS such as styleGAN2 by NVIDIA appeared in github and opened the way to faster training and more convincing results. It was then possible to create works of art which were impossible to distinguish from art made entirely by humans. More than than, it was possible to replicate the style of a specific art movement and create new works in that style (for example using the technique of style transfer published by Gatys et al back in 2015 “A Neural Algorithm of Artistic Style”)
Those new GAN structures work by compressing the information present in the dataset used to train them into the so called latent space. The latent space is a mathematical concept describing a space of (usually) 512 dimensions which is mapped to the space of images using the GAN. In simple words, the GAN allows to map one point in the latent space (512 coordinates) to a point in the image space (which is a representation of the work of art)
This mapping usually has the property that points near each other in the latent space also map to similar artistic images. Art creation with GANs now becomes akin to exploring the latent space and checking how the GAN translates each point in the latent space into a work of art. As such, AIart as also been called “latentism”. One interesting fact is that since this 512-dimensional latent space is very large, the mapping created with the GAN allows to create an almost infinity of art works. So in theory one could train the GAN on landscape paintings and have it produce and infinity of new original landscape paintings afterwards.
2- A new way to create art with technology?
While working on the hashers.ai project, I have talked to many artists who asked me whether they should be worried about this new technology and whether it will replace them in the near future.
As it stands, the A.I. is just a tool that artists can use to create art in a new way. One important aspect of AIart is that once the algorithm has been trained, it is then able to generate an infinity of new original works. This has important implications. Due to the mathematical nature of the approach, the A.I. is able, in theory, to surpass the human mind in terms of the sheer amount of works it can create. A normal human artist may be able to draw a new piece every week, whereas the AI can produce thousands per minute. That is not to say that each and every one of them will be compelling, but it opens the way to a new approach to creativity: Exploratory art creation.
Rather than having the design of the final work in mind, the idea is to let the AI create thousands of suggestions for the artist, and the artist is ultimately the curator and selecting the ones which are most in line with her aesthetic requirements and ambitions. It is possible that the artist will select only one piece among thousands created by the AI because he feels it best represents the vibe, the mood or the design she envisages. It is the artist’s function to change the many available parameters during the training process so that the output of the algorithmic pipeline matches her desired goals.
Now art creation becomes a back and forth game between the algorithm and the artist. The algorithm is suggesting possible works to the artist which selects her favorite ones and tweaks the algorithm further so that it produces works in line with what she wants. The artist doesn’t have a definite idea of the final result, just an idea of how it should look like. The AI does the heavy lifting of coming up with the final image or design because it has this capacity to create an infinity of works.
It appears the A.I. is a very powerful tool to enhance the artist’s imagination. It can propose thousands and thousands of new works, all unique, all original. At the same time, the DNA of the work is made up of the tweaks added by the artist to the pipeline, so ultimately it comes from her actions and her instructions.
Are there limits to this new approach? Yes, there are actually several limits which ensure there is still room for the human in the loop. First of all, while the scope of creativity is infinite, it is still limited to some extent to what the GAN has seen during the training phase. For example, if the GAN has never seen a car during training while learning to draw landscapes, it is unlikely it will start adding cars to the landscapes it will come up with. If the artist wants the AI to draw cars in the landscape, it needs to include some in the training dataset.
Another limitation is that it is not really possible to have the AI create a representation for some historical event. For now it would not be possible to have it imagine the storming of the Bastille in Paris on a summer day in 1789. Large compositions with many different figures are difficult to achieve for the time being, the AI still has difficulties to relate different objects in a picture and achieve perfect spatial coherence, but there is good progress in this area.
2- Is there a software to do A.I. art?
Another question almost all artists asked me is whether there is a commercial software which allows to create AI art. At the moment, there is no such software and the creation of art with AI requires to write python code and use deep learning libraries such as tensorflow or pytorch. One further difficulty is that training GANs is very compute intensive and takes a long time. To train a model able to generate good portraits may take up to a month on a x4 GPU stations using the latest commercially available GPUs. Making AI art is expensive for the time being and requires coding skills and understanding the neural network architectures used. While some applications for mobile phones allow to implement simple style transfer, they are only very basic. It seems that to get started with this approach artists will need to take up programming. It is very likely that in the future Python101 will be part of the curriculum of every decent art faculty.
3- Art | A.I. |Blockchain → Hashers.ai
Hashers.ai has worked with leaders in AI-art (among which D’AgostinoAI and IAmKaori) to bring the first ever large scale NFT collection of art created with artificial intelligence. We showcase what AI art can do, namely create an infinity of original and unique works, which is impossible to do by hand for a single artist.
A very successful recent NFT project , Hashmasks, created art by using a combinatorial approach. They made 15000+ art works by combining backgrounds, masks, objects and body types. The limitation of this approach is that the works are somewhat repetitive.
The collection we release is made up of 11,111 completely different artworks. We demonstrate how AI can solve this issue of repetitiveness but most of all we are proud of creating art we love to the core. I hope you enjoy collecting Hashers when the collection drops on 2nd June 2021.
If you wish to continue to learn more about AI art, you are welcome to Join our discord: Discord link. We hold discussions about how to create great AIart and discuss creation techniques.
Hashers website: Hashers
Instagram: @hashers.ai
Twitter: @hashersai
