Here's a demo I made:

Perform zkML image classification that can be verified onchain. No ZK circuit design needed—just Rust!

Use cases: perceptual rarity scoring for NFTs, proving branding/watermark/style, NSWF filters, visual access gating, airdrops to dog PFPs only, etc.

Here's how it works:

A program written in plain Rust, targeting the SP1 zkVM (succinct.xyz), receives an input JPEG and outputs:

- A class prediction based on a simple MLP

- An IPFS content identifier (CID), which acts as both a serving URI and an image content commitment.

runloop.xyz • memo.ai (acquired by Coinbase), summify.com (acquired by Twitter) • 🇺🇸🇷🇴🇨🇦🇨🇭

Source code, implementation details, and limitations: https://github.com/cgst/zkml-inference-sp1

The toy model used here is inadequate for real-world applications, but hey, it's just a demo!

See my recap of Frontiers https://x.com/gakonst/status/1826337591647785020!

In follow-on events we plan to leverage more of FC's crypto-native features, e.g. would have loved to have our Cursive ZKPs of connecting with people be mintable on the timeline, or re-use the Farcaster social graph in fun ways, or hacker-specific group chats.

More as we iterate...See you next year!

I started this demo at the Frontiers hackathon (https://warpcast.com/gakonst/0x584ab709) but only managed to clean it up this weekend.

Thanks to Succinct Labs and @uma for their support, and to @gakonst for organizing the awesome conference and hackathon!

The Succinct prover generates a proof of the execution that's small/cheap to verify onchain by an EVM app.

This proves that an IPFS image (e.g., an NFT's tokenURI at ipfs://...) has a specific class prediction according to our model.