The collection, titled Samba and hosted on sambacrypto.art, comes in 10,130 unique AI-based artworks inspired by the Brazilian Carnival. Easily the world’s biggest AIART NFTcollection, the Samba series is more than a group of fantasy elements; It’s a competition with its own rules. Yaak made a process based on a combination of Deep Learning (Artificial Intelligence), Machine Learning, and Data Science to create this competition.
Yaak has worked with GANs to create the images, and other AIs that act as judges to evaluate and rank the artworks, also identify the rare items – scenes that are impossible in the real Carnival like two Masters of Ceremonies dancing.
Let’s hear more from the artist himself!
The collection is inspired by the Carnival in Rio. What makes it so special for you?
Carnival is culturally deep-rooted to Cariocas (residents of the City of Rio de Janeiro, Brazil); Sometimes, we say that the year only begins after the Carnival. The world remembers the Carnival as a colorful and festive party. However, it is, in essence, a competition with lots of rules, evaluations, and time frames. Each year, by the end of the Festival on ash Wednesday, we have a live tv show presenting all Samba schools notes; the notes will determine the position and privileges of each school in the next year. This math behind the party fascinates me, so the Samba Collection is an essay that tries to simulate this math of aesthetics evaluation through Artificial Intelligence use.
One of Yaak’s favourite pieces, Samba #000365
How does one go about creating a collectible series?
I’m interested in this space in between collectible and AI art. Samba is at the same time a collectible series that has its own discovered rarity levels summarized in the card. Still, also Samba is AI Art Series, with the original GAN-created artwork attached in each item by an IPFS link.
The interesting point in my view is that collectible rarity and art appraisal are not necessarily directly related. The Sambas I most like, are sometimes those identified by the intelligence as the items with fewer scores, and maybe the cheapest ones.
To create the AI Artworks, I started compiling the dataset by selecting thousands of Carnival images, then by a computer vision approach, and I kept those that show the Couple of dancers. Next, I used an algorithm to identify the frames that best represent movements of the Couple dance. At this moment, all the initial images were discarded, and about five thousand movement-blurred frames were used to train the Neural Network that created the AI artworks.
If this was just an AI series, I could Stop here. But to simulate the competition aspects and find the treasures in the latent space, it was necessary to go deep. So Each image was evaluated by a computer vision algorithm that tries to identify and score each object in the scene, acting like a carnival judge. A statistic analysis identified five groups of average scores; then, I named these groups using the same typology of the Carnival: Special, Gold, Silver, Bronze, and access. As in Carnival, the evaluation determined which series each item belong to.
To create the AI Blocks or Samba Schools, I took advantage of another deep learning method that ended by group images in clusters due to similarities in the features of the pictures. I named different levels of the clusters as places in Brazil and samba-related expressions. So the Name of each AI Block is a code to identify the two deeps of hierarchical clustering.
One of the top scored Samba
What is the main difference between 10K and 130 rare items? In simple terms?
Each Samba has unique characteristics, but the 130 Rare have a combination of elements that are totally imagined by the AI (121 Scenes) or are in the top-scored components by the AI algorithm (3 top Flag-bearer, 3 Top Masters, and 3 top Couples).
The rare combinations found that is not possible in the Carnival are:
Two Masters in the same scene, a Flag-bearer with two flags or a Master holding the flag. These combinations are not in the dataset, actually are not allowed in the real carnival rules, but in the AI dimension, this phenomenon emerged.
A Fun fact: First, the AI created the 10130 items; the number is arbitrary. There is no magic here. The fact We ended with 130 Rare elements was just a coincidence; math sometimes makes jokes like that, and here is the magic.
Detail of Samba #002499
People in NFT often use the term “generative art” in relation to collectible projects as opposed to artists using GANs. As an artist that used both – are there similarities in the approach between the two?
As in the Generative Art collectibles, Samba has Rarity levels and Attributes. But the approach to get the rarities is different. It’s less deterministic and more exploratory.
In Generative Art Collectibles, usually, the rarity is created by combining attributes algorithmically; the rule is in the code, it is pre-determined. The approach of Samba is the opposite; there are no rules around the elements being created, the Neural Network learns to generate new images statistically similar to a data set. So for the GAN, the images produced are the best average that matches the Data set, or in other words, the network is trying not to create anomalies. So to identify rarity in this latent space, I had to develop a process combining machine learning, computer vision, and statistics to group elements by similarity, identify components, evaluate and classify, ending by finding scenes very rare.
Before the process started, I had no idea what kind of image or combination will be created by the GAN, neither that the GAN will output things that are not in the original frames or how many times this will happen. The details in cards are my attempt to highlight these phenomenons, but yeah, this turns the items collectibles, as there is a sense of value around rarity.
You can learn more about Samba via https://sambacrypto.art