No, I'm not confusing it. I realize that LLMs sometimes connect with diffusion models to produce images. I'm talking about language models actually describing pixel data of the image.