A brief overview of neural networks that have been uploaded to the blockchain, and why it matters for NFT artists and collectors.
Challenge: On-chain AI NFTs
Consider an NFT that wants to robustly store its metadata on-chain, but it also uses an AI model, specifically a neural network, to generate outputs.
AI image models today typically consist of billions of parameters, resulting in models that are tens of gigabytes in size. Blockchains are limited to upload sizes on the order of ~hundreds of kilobytes, even when using technologies like Art Blocks BytecodeStorage with on-chain compression.
Therefore, any fully on-chain AI-based NFT must have an extremely compact model with as few parameters as possible. It must balance output complexity with the number of parameters, and be a custom, artist-designed model.
History of on-chain AI NFTs
note: There are tens of thousands of NFT projects on Ethereum, so I could be missing projects in this history – please reach out if you are aware of any projects that should be added to this timeline
The following projects are the only ones that I am aware of that store their AI model entirely on the blockchain. They all consist of purpose-built models designed from scratch.
XOR by Nahiko
XOR claims the title of “first on-chain neural network”, with a single XOR gate that operates on input controlled by the artist and collector.
Aragnation
by Devi Parikh and Abhishek Das
This generative art project uses a small multilayer perceptron to predict the saturation of colors in some of the pieces.
InChainPepeGAN by @LifeOnChain_NFT
A fully in-chain neural network is executed inside Solidity to generate 5×5 outputs, with 3 frames. This project aimed to produce the NFTs svg outputs while remaining under 30M gas.
I. by ryley-o.eth
A hyper-compact, quantized cVAE-GAN is used to translates generative expressions to 64×64 outputs that look like pixelated photographs of human eyes.
What’s Next?
I enjoy this game and will continue exploring how custom designed AI models can be used as a tool to generate compelling outputs within the strict storage constraints imposed by blockchain networks.
I’m currently researching some novel model architectures that I’m excited to use in future projects.
In theory, custom neural networks may actually represent best compression algorithms, but they require expertise, experience, and a hardware to design and train. Despite on-chain neural networks currently being extremely rare in the NFT landscape, they may prove to be a great fit for on-chain blockchain art in the long term!
For more information about my project, I., see my blog post below!
NFT AI Models on the Blockchain
A brief overview of neural networks that have been uploaded to the blockchain, and why…
I.
https://www.artblocks.io/collection/i.-by-ryley-o.eth An on-chain neural network generating expressions. I’ve had this itch for years to see…
Blockchain Value Propositions for NFTs
Diamond Contracts and Diamond Storage
Blockchain Primitive: Delegation Registry
Leave a Reply