Author: Josh O'Sullivan, CoinTelegraph; Translated by: Baishui, Golden Finance
Ethereum co-founder Vitalik Buterin has endorsed the new Token for Image Tokenizer (TiTok) compression method for potential blockchain applications.
Not to be confused with social media platform TikTok, the new TiTok compression method significantly reduces image size, making it more suitable for storage on the blockchain.
Buterin highlighted TiTok’s blockchain potential on decentralized social media platform Farcaster, saying “320 bits is basically a hash. Small enough to put every user on-chain.”
This development could have significant implications for digital image storage for personal profile pictures (PFPs) and non-fungible tokens (NFTs).
TiTok Image Compression
TiTok, developed by ByteDance and researchers from the Technical University of Munich, can compress images into 32 small data blocks (bits) without losing quality.
Advanced artificial intelligence (AI) image compression enables TiTok to compress a 256x256 pixel image into “32 discrete markers,” according to a TiTok research paper.
TiTok is a one-dimensional (1D) image tokenization framework that “breaks the grid constraints present in 2D tokenization methods,” resulting in more flexible and compact images.
“As a result, it can significantly speed up the sampling process (e.g., 410 times faster than DiT-XL/2) while achieving competitive generation quality.”
TikTok research paper showing a comparison of image compression sizes. Source: TikTok
Machine Learning Images
TiTok leverages machine learning and advanced AI to convert images into tokenized representations using a transformer-based model.
The method uses regional redundancy, which means it identifies and uses redundant information in different areas of the image to reduce the overall data size of the final product.
"Recent advances in generative models have highlighted the important role of image labeling for efficient synthesis of high-resolution images."
According to the research paper, TiTok’s “compact latent representation” can produce “more efficient and effective representations than traditional techniques.”
Illustration of image reconstruction (a) and generation (b) using the TiTok framework (c). Source: TiTok
TikTok, not TikTok
Despite the similar name, social media platform TikTok has not been endorsed by Buterin.
The Ethereum co-founder highlighted TiTok’s blockchain potential, lending credibility to this new AI-driven approach to image compression.
"Different from existing 2D VQ models that view the image latent space as a 2D grid, we provide a more compact formulation to label images as a 1D latent sequence."
The proposed new method can "represent images with 8 to 64 times fewer markers than "2D markers," and the team hopes that this research will shed light on "more efficient image representation."
