We have Deep Learned Gobardhan Ash's Avatar series. These are works created by the Artificial Intelligence engine. Each image of 48x48 pixels has been inscribed on the Bitcoin blockchain. The inscribed image itself is distinct and precise. In line with the ethos of Ordinals they do not need an off-chain image for clarity. The images have also been compressed to less than 1 kb allowing for minimal transaction costs. And keeping size to a minimum we have decided not to include any metadata - that would add another 50 bytes at a minimum and in our case only add the name, an index number which is of no consequence. Note that there are no trait names, as each art work is unique. Some of the traits may look similar but are all different.
Earlier Sales
The original paper on GAN's was published in 2014 by Goodfellow et al. The computing power was still not there. It was only around 2017 that we began to see some interesting experiments using GAN's.
It should be noted that Neural Networks or Convolutions are not new. The author himself having coded up convolutions in 1997 [1] and having attended a class on Neural Nets & Fuzzy Logic by Lofti Zadeh - where the avant-garde use-case was to identify when clothes in a washing machine were clean enough! The advent of GPU's changed everything as Neural Nets are inherently "parallelizable".
AI Generated Nude Portrait #7 Frame #64 - Sold for 842,116 USD Sothebys London 2022 (Image Courtesy: Sothebys)
AI Generated Portrait of Edmond de Belamy - Sold for 432,500 USD Christies 2018 (Image Courtesy: Christies)
Generative Adversarial Networks
These have evolved! since 2018 and in fact, the AI tech is moving so fast that six months seems to be five years. GPU's are bursting exponentially, state of the art is continuously changing. Though the focus of a lot of these technological battles are specifically targeted at LLM's where there seems to be a desire to have as much data coupled with as much computing as possible. Most GAN's have been trained with close to 100,000 images possibly supervised. This also highlights a problem, what if we have fewer data points, and much fewer images than what is expected in training a GAN. Style transfer on top of a pre-trained pickle file does not really work for art - the images created are too far from the artist's own creation. Using FFHQ or Huggingfaces pickle files and then attempting a style transfer is not appropriate. As curators, we have detailed knowledge/understanding of the artworks from which we are training. We have been able to "solve" the problem by repetitive augmentation, by standardization, with the knowledge of the compositional pieces, with the knowledge of what's relevant, the ethos of the artist, by breaking and joining, by mixing and matching the compositions. We feel this is a superior way of creating profile picture projects. We realize that this is a "categorization" approach to AI
To give you an idea of the complexity, please see below the FID graph which converges close to 10 before bouncing up never to seemingly fall back no matter how much we try. The FID score achieved is "not bad" given that the best achieved by GAN's being trained from the FFHQ database with much higher resolution images reaches close to an FID of 5 ; however, these are trained with 70,000 plus images. Note we (at Prinseps) are dealing with fine art, not very high-resolution, and in this case with a rich set of colours. We feel the results are quite encouraging. More importantly, as an experiment in modern art, we find the generated images to be both modernist and appropriate. These are certainly very different than the cartoonish images that proliferate the NFT / Ordinal space that certainly don't qualify as modernist. The AI engine is creating legitimate newer artworks ! This is a tipping point in Art/Image generation. This is the moment the 'SkyNet' switches turned on !
Some of the images may be close (in the neural net), and ideally, we should use a measure to select images that are the furthest from each other. But this is not an easy problem. There is a concept of an Euclidian distance; however, that is suited for black and white pixels. The artist was a master colorist - he was able to achieve texturing in a two dimensional space by using water-colour what would probably have been more easily done in oils. He was able to create depth that would generally only have been possible in an impasto art work executed in oils. We are enthused that the StyleGAN engine is able to learn and generate the richness of the colours and the depth in execution. This is certainly encouraging for future projects. The color palette in these works is high and they cannot be minimized to black and white for a euclidean distance calculation. Given an initial image, it is somewhat of a reverse optimization !to find a "far" image. This is still up for research.
It was not easy, we have failed on many attempts and produced the following images with exponentially high and ridiculous frechet scores. The trick is to add structure to the training which is very much at odds at how the LLM's have been trained with as much data as possible. Sometimes less is better. Our implementation interestingly mirrors exactly the new research which is being published elsewhere.
To finally achieve success
The GAN-generated images display the latent walk. We focus on a feature - a specific background. And compare the difference just of the backgrounds and then the difference of the difference to show the "walk". Since each artwork is different it would not be possible to use a recursive implementation of an ordinal. However at approximately 350 bytes for a 24x24 pixel image which can be distinguished, the transaction cost per ordinal will be minimal..jpg)
.jpg)
.jpg)
The same "walk" is visible in the faces. We focus on a specific face type.
However, note for a handful of images we still see failure. Here are four samples as such with the last of these showing double pupils. We have decided to go ahead and include these in the Ordinals release so as not to cherry-pick.
We believe this is cutting-edge. But more importantly legitimate modern art with newer artworks created in the artist's style. Some of the artworks are below. We do hope you find them aesthetically appealing.
References
Chatting with GPT4.0 about 'The First AI vatars' and other related projects
The Revolutionary Wall of Artificial Intelligence
"Analytical Models for Vehicle/Gap Distribution on Automated Highway Systems", H.-S.J. Tsao, R.W. Hall, and Indrajit Chatterjee, Transportation Science, Vol. 31, No. 1., pp. 18-33, 1997.
Position: Categorical Deep Learning is an Algebraic Theory of All Architectures, Gavranovic´, Lusard et al
Numerically Ranking The ArtWorks
The SkyNet moment - GPT4.0 is NOT able to recognise AI-generated artworks
A Short Presentation at the Soho House, Mumbai on May 22, World Bitcoin Pizza Day
A Panel Discussion at Club Jolie's Worli On June 19th, Discussing Bitcoins & Ordinals
