Why would you feed pixels into transformers? Surely you could translate the overall image into frequencies using Fourier transforms and then into embeddings which feed the LLM?