How Krea trained its image model's prompt expander with RL

Krea trains a language model to turn short image requests into detailed descriptions its image generator can use. Its Krea 2 technical report explains that users' brief prompts differ from the rich captions used in image-model training, and the prompt expander is meant to bridge that gap while keeping the request's meaning.

Training starts with paired short and detailed captions. Reinforcement learning then optimizes the expander through the images it produces: rewards cover image quality and preference and whether the expansion stays faithful to the original request, and a diversity reward discourages collapse into one visual style. Krea says it used Thinking Machines' Tinker to prototype reward models and train the expander, per its technical report and a July 2 post. The report gives no quantitative results for the expander, so this explains the design, not how well it works.