An #image needs 40 bytes instead of 196,608
source: https://arxiv.org/pdf/2406.07550
The Chinese company #Bytedance, which includes #TikTok, has developed a series of transform models for image #generation and #compression with up to 307 million parameters in collaboration with the University of Munich. It can reconstruct a square image with a length and width of 256 pixels largely correctly with just 32 tokens. That is 40 bytes of the original 196,608 bytes. The model can be used both for compressing image data and for pure image generation and is said to be up to 410 times faster than conventional diffusion models.
