My approach is very different. It uses an image model that has multimodal support. The links are extracted from the generated image and rendered based on context. It's very different and mostly has never been done with an image model before due to the lack of multimodality.

→ View original post on X — @skirano,