The rise in interest in NFTs is related to, and also driven by tools that allow the programmatic production of massive amounts of visually pleasing art and themed galleries. Tools like HashLips and software packages for R and Python allow both coders and non-coders to create thousands of themed images, some of which can do quite successfully on the NFT market.
But the ability to mass-create thousands of highly similar artistic outputs, which oftentimes are close variations on a theme, poses a set of new and inter-related challenges for graphic-creators wishing to profit from those on the NFT market.
The core cause is that more choices are not always better, and particularly when it comes to NFT galleries, ‘more’ creates three very different types of problems: information overload, paradox of choice, and product cannibalization.
Information overload is intuitive: if a creator produces 500 graphical variants of their favorite, say, sombrero-wearing mushroom, simply viewing and absorbing this massive amount of information while browsing a gallery would be an enormous challenge.
The paradox of choice is also related to the availability of multiple choices but is more difficult to counter because it occurs even when there is no information overload at all. It captures the fact that people find it easier to choose (and in fact, spend money) when there are fewer choices than when there are more choices. When setting up an NFT gallery, a creator will want to plan in advance, limit the options to only the best ones, and avoid unnecessary options.
Product cannibalization refers to the situation where a seller loses revenue because they introduce a choice that lowers demand for an existing choice in the same cohort/portfolio. In the context of an NFT gallery, an item that would otherwise appear to stand out in terms of creative aspect, could be cannibalized in value by introducing another one that is highly similar.
Here we show how AI systems, and specifically, biometrically-informed AI can offer effective solutions for these core problems. These systems allow advance planning of optimization strategies for constructing the set (assortment) of NFTs.
The approach we present, based on Aifilia technology, begins by first understanding the ‘similarity space’ of a given set of NFT images. Informally, this measures the similarity of each of the images to each of the other images, producing what is called a similarity matrix. Examining this similarity matrix can indicate which (if any) image particularly stands out, which images are generally too similar to other ones, or even if there are ‘clusters’ of highly similar images.
We have not seen these tools being publicly applied to analysis of NFT ‘real estate’, but it is safe to assume they already are. The main problem is that, even when using cutting-edge AI tools, the way that AI determines the similarity between images doesn’t necessarily emulate people’s perceptions.
Aifilia’s biometrically-informed AI directly addresses this weakness. To show the advantages of biometrically-trained AI we will analyze an assortment of seven image NFTs, which we've been permitted to use for this test. As can be seen, of the seven images, five (1 to 5) are non-abstract graphics capturing human themes, and two (6 and 7) are abstract graphics generated programmatically.
tl;dr: A cutting-edge Deep Neural Network for vision determines that images 6 and 7 are highly similar, as are images 3 and 4. However, it fails to discriminate the abstract images from the non-abstract images. Notably, it determines that images 2, 3, 4 are all more similar to images 6, 7 than to image 1.
We first analyzed these images using a modern AI system. We did this by feeding the images into a convolutional neural network (VGG-19), extracting the resulting activations (image-embeddings) from the network, and quantifying image-similarity based on the cross-correlation of the image embeddings. The resulting similarity matrix is presented below, and shows good signs of sanity/validity: the two mushroom images (3 and 4) are rated as highly similar, as are the two abstract graphics. In addition, image 2 is more similar to images 3 and 4 than to the others.
Still, there are some problems with the AI assessment.
The main limitation is the following: looking at images 1 to 5 (rows 1-5 in the similarity matrix), we see that for each of those, the similarity with the abstract shapes (6 or 7) exceeds the similarity with at least one of the non-abstract images. This means that the abstract images are not identified as ‘different’ by the AI engine.
In fact, this AI indicates that images 2, 3, 4 are generally more similar to the abstract shapes 6, 7 than they are to image 1.
tl;dr: Aifilia’s biometrically-trained AI produces substantially more accurate information about the perceived similarity of the images in a way that matches human intuition. Its conclusions include all the important ones made by cutting edge AI, but it also decisively and consistently separates abstract from non-abstract images.
Aifilia’s biometrically-informed AI also constructs a representation for each image, but through a proprietary process that is informed by a mathematical model that can predict people’s biometric responses to those images (note: we don’t measure responses to these images, but predict them). We used these representations to produce a similarity matrix from Aifilia’s knowledge.
As seen in the matrix below, the technology provides higher quality information that subsumes all that derived from the cutting edge AI, but reflecting human intuition much better. Specifically, it also shows high similarity between the two mushroom images, as well as between the two abstract images.
However, in a marked departure from the off-the-shelf AI system, Aifilia’s analysis indicates a much weaker relationship between the non-abstract images (1 to 5) and the abstract ones (6 and 7). In fact, in Aifilia’s analysis, all five non-abstract images show a consistent pattern: each shows a stronger similarity with all other non-abstract images than with the two abstract images.
This is a massive difference from the conclusions of the cutting-edge AI, which did not even identify one (not to mention five) non-abstract images showing such a strong differentiation pattern.
NFTs are big business now and will become a huge business in the near future, once the hype of ‘minting’ events fades, and NFTs become a standard tool for IP management and valorization. However, the ability to generate hundreds or thousands of images programmatically presents new challenges, which increase uncertainty both on the side of sellers and purchasers.
We presented here several use-case scenarios where AI can be applied to evaluate a set of NFTs. We showed that AI can help determine the relative uniqueness of each NFT image in a set, and whether there exist sets of highly similar graphics. This advance-knowledge is crucial to optimization of content for galleries, and for dynamic adjustment in gallery content through the sales process.
While a cutting-edge AI architecture could already provide useful observations, it also made several invalid predictions that do not align with intuition. In contrast, a biometrically-trained AI does much better because it knows which dimensions of the images are important for humans.
If you want to be updated on these issues and Aifilia's NFT applications, you may sign up below, or you may follow us on Medium and Twitter.