I’ve cleaned up my LLM prompt-enhancement script and added it to the repo. The latest improvement to my partially-enhanced prompting was to add markers to the input so that the LLM would only see the part of the dynamic prompt that I wanted to enhance:

${a=!__artstyle__} ${a} __pinupgal__ @< ${a} __background__ >@

The “!” forces the variable to be evaluated immediately, so that the same value is used in both places. I’m including the style string in what’s passed to the LLM so that it doesn’t end up contradicting the style used for the gal. With modern models that prefer detailed prompts, the potential repetition doesn’t hurt, and might even help.

This is significantly improving the hit rate when I deathmatch the results. I’m still tinkering with the system prompts and the LLM settings, because not only is there a tendency to add random captions to the picture, sometimes it gets much too creative.

Getting started…

---

Tinkering with multi-gal…

Romancing the 1girl…

---

Comments via Isso

Markdown formatting and simple HTML accepted.

Sometimes you have to double-click to enter text in the form (interaction between Isso and Bootstrap?). Tab is more reliable.

---