Pick a category, roll the dice, type nothing. It kept producing the same pose until I found I was feeding old prompts back into the context.
About the project
This one exists purely for fun. I am fascinated by how these models think, and I wanted to see how far off the rails I could push an image generator without typing anything. So I put a language model in front of it. You pick a category, roll the dice, and the model hallucinates an abstract prompt before any image rendering starts.
Early on it fell into a rut. Every character came back in the same pose, run after run. It turned out I was accidentally feeding the previous prompts back into the new context, so the model was quietly agreeing with itself. I isolated the client-side state so the app forgets every old generation between runs. Even with a clean slate, individual models drift toward their own habits, so I added a switch to route prompts across different versions and break the convergence.
The whole session history lives in the browser. I roll the dice, see what comes out, and roll again.
- React
- TypeScript
- LLM Integration
Gallery
