Thank you for your helpful comments. I'm working on a browser version to address the first point. Unfortunately, the Highly variable images are an intrinsic limitation of pulling images from gelbooru based on tags--the gelbooru tagging system is very flat so there is no way to connect tags to characters, leading to tag bleed across characters. There also may simply be no image that hits all the tags. I've set it up to prioritize the sex position over character traits in this case. My longterm goal is to have a set of pre-baked scenes and poses the llm choses from based on the narration, then insert characters in, but this is substantially more work. The gelbooru images are a placeholder.