A novel technique for effective onchain image compression called TiTok AI may prove beneficial for blockchain applications. The new Token for Image Tokenizer (TiTok) compression technique gets the support of Ethereum co-founder Vitalik Buterin due to its potential blockchain application. The new TiTok compression technique, which is not to be confused with the social media network TikTok, greatly reduces image size, making it more feasible for blockchain storage.
On the decentralised social media network Farcaster, Buterin emphasised TiTok’s blockchain potential by saying, “320 bits is basically a hash.” Minimal enough to fit on each user’s chain. The advancement may have noteworthy consequences for the digital image preservation of non-fungible tokens (NFTs) and profile photographs (PFPs).
TiTok, created by academics at Technical University Munich and ByteDance, enables the lossless compression of an image into 32 tiny data units (bits). The TiTok research article claims that TiTok can compress a 256×256 pixel image into “32 discrete tokens” thanks to sophisticated artificial intelligence (AI) image compression.
TiTok is a framework for tokenizing 1-dimensional (1D) images that “breaks grid constraints existing in 2D tokenization methods,” producing images that are more compact and adaptable. “Therefore, it results in a significant acceleration of the sampling process (410 × faster than DiT-XL/2) while achieving a generation quality that is competitive.”
TiTok uses sophisticated AI and machine learning to transform photos into tokenized representations using transformer-based models.
The technique employs region redundancy, which finds and utilises redundant information in various image regions to minimise the final product’s total data size. “Image tokenization plays a critical role in the effective synthesis of high-resolution images, as demonstrated by recent developments in generative models.” TiTok’s “compact latent representation” is able to produce “substantially more efficient and effective representations than conventional techniques,” according to the research paper.
Buterin did not support the social media app TikTok, despite the name being identical. The co-founder of Ethereum gives credence to the innovative AI-driven image compression technique by showcasing TiTok’s blockchain potential. We offer a more condensed formulation to tokenize an image into a 1D latent sequence, in contrast to the current 2D VQ models, which treat the image latent space as a 2D grid. The team expects the research might shed light on “more efficient image representation,” as the proposed new method can “represent an image with 8 to 64 times” fewer tokens than “2D tokenizers.”