Reinforcement Learning for Code-Based Image Generation

Researchers trained a language model to generate editable images by writing code via reinforcement learning. The method addresses the lack of granular control in prompt-based generation by using a judge-based reward function to evaluate aesthetic quality.