tl;dr: Researchers trained Qwen 3.5 35B via reinforcement learning (GRPO) to generate images by writing p5.brush JavaScript code, making the artifact editable rather than locked behind prompts. The main challenges were designing a reward function for subjective aesthetics—solved with pairwise judging against a hand-rated pool of 581 reference paintings—and preventing API hallucinations, which was fixed by replacing a 400-line API reference with a short allowlist evolved via GEPA. The authors concede it's slower than standard image generation but argue it opens up finer-grained creative control.