The sad thing is that this has been possible for decades using convolutional autoencoders, but with LLMs we forgot that AI architectures other than transformers still exist.