RL-trained Qwen painter
Technique · Image
Qwen fine-tuned with GRPO reinforcement learning to write painting code, with a playable demo; no weights or repo linked
StatusDemo
Evidence
- Aug 23Reward mixes compile gate, length, HPSv3 aesthetic score and a pairwise judge; pairwise rubric plateaued about 3x faster, output shrank from ~13,500 to under 2,000 tokens
- Aug 23Commenter says they did similar RL on SVG to draw pelicans, and that most behavior came from SFT; RL is inefficient at style or novelty out of distribution
Alternatives
Works with