Image Encoding - ueaj - Obsidian Publish

We need variable length image tokenization. I feel like the fixed encoding paradigm is very anti-information-theory pilled. The issue with current approaches is that they either have enough bandwidth to encode the full image, in which case it takes up a ton of compute, or they compress the image to a fixed size, in which case it falls apart on edge cases. Either you My model for why this is the case is information-theoretic, our tokenizer is the transmitter and the signal is the tokens, our transformer is the receiver and it's destination is whatever helps in NTP. Of course the set of real world images is vastly smaller in size than the set of images encodable in the pure RGB signal, this is why we have compression algorithms like JPEG and PNG that take advantage of common artifacts in real world images. This opens up the possibility of having an architecture or tokenizer that sends just the necessary amount of information into the transformer that encode the image, rather than the ent

saved by

• Ali Ahmad