Tria Chromatica (meaning: three produced by colour) pays hommage to 20th-century contemporary artists by combining their creative talent with modern technology. Inspired by some of the world’s greatest geniuses, Tria Chromatica consists of three unique, AI-generated collections; ‘Stimaes’, ‘Swayils’ and ‘Limah’.
Stimaes is the first, mother collection — and features an AI that has been studying three pieces (The Snail, Memory of Oceania & The Codomas) of the artist Henri Matisse.
STIMAES
The second collection in the Tria Chromatica series is Swayils, based on the revolutionary abstract artist Wassily Kandinsky. Swayils is an AI’s interpretation of Wassily’s works of the Composition II (1923), Orange (1923), Composition VIII (1923) and Untitled (1934).
SWAYILS
The third collection of Tria Chromatica, named ‘Limah’, is based on the works of the mystic Hilma af Klint. The Swedish artist is credited with creating some of the earliest known abstract paintings in Western culture. Her work precedes both Kandinsky and Mondrian’s first totally abstract compositions.
LIMAH
When a group of artists came together to talk about art legends, history and creativity, Tria Chromatica was born. As an artist collective, we began to question how great artists like Matisse, Kandinsky and Klint’s work might have evolved and progressed if they hadn’t passed away. As the conversation progressed, the subjects of modern art, algorithms, and artificial intelligence were eventually brought up. Can we, as humans, who are fortunate enough to live in an era of advanced technology, use superhuman intelligence to feed data from the artist’s past to make an artist invincible? The artists featured in Tria Chromatica were all regarded as prominent personalities in the modern art scene and we would like to highlight these to this new generation. Would machine learning be a good way to indirectly extend their body of work? Tria Chromatica was created to put this hypothesis to the test.
The difficulty wasn’t to create random art pieces using elements from Matisse’s, Kandinsky’s and Klint’s works. If Tria wanted to pay a respectful and modern hommage to these great artists, it needed to develop stand-out imagery with a strong aesthetic appeal. As a consequence, Tria opted to concentrate on pieces by each artist that work well together. It was also integral that when creating these pieces, collectors would feel as though the artist could have indeed created them themselves.
Building the Tria Collections
Tria Chromatica had to use deep neural networks to replicate, recreate, and blend styles of artwork when creating the AI art. This was accomplished by teaching the AI to understand existing pieces of art. (In this scenario, Matisse, Kandinsky and Klint were the subjects.)
GANs (generative adversarial networks) are computational structures that pit two neural networks against one other (thus the name “adversarial”) to produce fresh, synthetic examples of data that can pass for real data. When compared to regular GANs, StyleGAN2 is a generative adversarial network that enhances image quality and improves image augmentation.
Therefore, Tria Chromatica decided to leverage this form of network to create high-resolution abstract art from a sequence of training images that were fed into the network. The generative adversarial network training required that Tria first turn all of its input data into a single dimension and then build an extremely precise structure around it.
The paucity of data was a major issue as the learnings were based on a small number of pieces for all three collections. A lot of data was required for the AI to curate outstanding art. The more data from the artist’s work that the AI can learn and analyse, the better the outcome. Therefore, individual pieces by Matisse, Kandinsky, and Hilma af Klint were very precisely isolated, altered in colour, size, opacity, transparency, and positioning to create a larger amount of data that could be supplied to the AI.
Although the input data increased, the data volume remained low. An adaptive discriminator augmentation mechanism (ADA) was adopted by Tria as a result. ADA helps StyleGAN’s learning to be more efficient, even with limited input data, and it does not require any major network design modifications.
One of the challenges Tria ran into after the training was that the output was still a little hazy and a little too abstract to properly honour the artists from whom we drew inspiration. Work in its earliest stages of development was blurry, ill-defined and lacking in the vibrant colours Matisse, Kandinsky and Klint are known for producing. It was so decided to use an ESRGAN (Enhanced Super-Resolution Generative Adversarial Network) -based network to improve textures and optimise the output. ESRGAN works by analysing an image’s pixels for colour variation, and then injecting the details it thinks should be there based on data it has previously seen.
After optimising the forms, patterns, and colours with StyleGAN2 and ESRGAN in post curation, the AI produced thousands of pieces of artworks, many of which still didn’t fit the inspired aesthetic. The final pieces of the entire collection, which now reside on the Ethereum blockchain, were handpicked and cleaned up from a big pool provided by the networks.
It took over 6 months to develop the initial basic IP and methodology for the mother collection.