Vision language models have processed images as raw input to neural networks since late 2023.