Fully Decentralized NFT Storage through Frequency Domain Neural Network Compression

Abstract

Non-fungible tokens (NFTs) and other Web3 multimedia are typically stored off-chain because image and video assets exceed practical on-chain storage budgets, creating a gap between immutable ownership records and long-term media availability. This paper proposes a collection-level neural compression approach that converts an entire NFT collection into a single compact decoder. The decoder maps a token's integer index directly to its reconstructed image, and is intentionally trained to memorize the collection so that no per-image latent codes are stored. To minimize the decoder's on-chain footprint, we parameterize weights in the frequency domain, progressively prune high-frequency coefficients via zigzag-ordered masking, and apply run-length plus Huffman entropy coding to the resulting sparse parameters. Across three 10,000-image NFT benchmarks, Bored Ape Yacht Club (BAYC), Azuki, and CryptoPunks, the compressed artifacts are reduced to 7.63 MB, 17.09 MB, and 3.33 MB, respectively. These artifacts achieve up to \(177\times\) smaller size than PNG while maintaining high reconstruction quality, measured by Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), with PSNR \(\geq\) 33 dB and SSIM \(\geq\) 0.967. We further report Ethereum storage gas estimates showing that collection-level on-chain deployment becomes feasible at \(10^{3}\) – \(10^{4}\) USD under representative fee assumptions, reducing on-chain deployment cost by 58.6%–99.4%. The main contribution is a collection-level compressed decoder that serves as the deployable on-chain storage artifact, removing dependence on external media hosts and avoiding per-image latent storage. These results demonstrate a practical path to self-contained, on-chain availability of large NFT multimedia collections without relying on external storage networks.