Comment by edg5000

I think using the vision decoder baked into modern LLMs is the way to go. Have the LLM iterate; make sure it can assert placement qualities and understands the hard requirements. I think it can be done.