I should correct myself a bit here. I believe it's actually UNET type models that can be used to very lossily "compress" images. You can use latents as the images, and use the locally running image generation model to expand the latents to full images.

In fact any kind of decoder model, including text models can use the same principle to lossily compress data. Of course, hallucination will be a thing...

Diffusion models, depending on the full architecture might not have smaller dimension layers that could be used for compression.

Just discard images and have the LLM describe them on the other side.