The Ethereum co-founder recently expressed interest in the TiTok AI, prompted by a post from a researcher at the artificial intelligence (AI) image generator platform Leonardo AI, who goes by the handle @Ethan_smith_20. The post briefly outlined how the method could be utilized to encode complex visuals into 32 tokens, benefiting those interested in reinterpreting high-frequency details within images. According to a research paper on the image tokenization method, TiTok employs AI to compress 256 by 256-pixel images into “32 distinct tokens.”
The paper highlighted shortcomings in previous image tokenization methods like VQGAN, which were constrained by using “2D latent grids with fixed downsampling factors.” These previous approaches were limited in their ability to handle redundancies in images and struggled with similarities in closely positioned regions. The new method is also significantly faster, boasting speeds up to 410 times faster than current technologies, a substantial advancement in computational efficiency.
