I have a version that does that for the image input models, I hope to write that up soon

→ View original post on X — @simonw