Training Qwen with RL to Paint Watercolors via Code

zsakib_ · x · 2026-08-24

Surya Narreddy demonstrated using Reinforcement Learning to fine-tune Qwen 3.5 35B to generate watercolor paintings by writing p5.js code, rather than generating images directly.

How it works:

The model receives a prompt and writes a complete p5.brush JavaScript sketch.
The code is rendered as a PNG in a sandboxed Puppeteer environment.
A judge model compares the output against hand-rated reference paintings.
The judgment is converted into a reward signal, updating the model via GRPO.

Reward Checks:

Code compilation status.
Code length.
HPSv3 preference scores (model-judged preference).
Human pairwise ratings from hand-picked sketches.

This approach makes the generated image editable code, addressing the limitation of traditional generative models where outputs cannot be directly modified.

Original post →

Training Qwen with RL to Paint Watercolors via Code

More from coding & agent