The vision model kept rendering muscle definition straight through the shirt.
About the project
The end goal is a fully generative tabletop RPG. I want a player to sit down, generate a character with as much or as little detail as they feel like putting in, and come away with a persona that is actually theirs. Character generation was the first piece I could build.
I started it in Gemini Canvas and the app rotted once the file hit a thousand lines. The interface just broke. I moved it to Cursor and split it into modules, and once it compiled, the model started fighting me on basic logic. The vision pipeline could not handle clothing layers. Describe muscle definition and it rendered bare skin straight through the shirt. I fixed the bleed by looping reference images back through the image model to enforce a consistent outfit.
It runs in two directions now. Dial in conditional attributes and it generates a structured profile with artwork, or feed it reference art and the vision model extracts the physical data. Everything saves in the browser. The next step is generating 3D models from the character art so players can print their miniatures and carry the same character into virtual play.
- React
- TypeScript
- LLM Integration