Reinforcement learning
Hugging Face trains coding agents to paint digital watercolours with TRL
Using Transformer Reinforcement Learning (TRL) and OpenEnv, researchers successfully trained a coding model to generate vector drawing instructions that paint watercolor-style images.
Original cover art, generated for this story. THE VISSION does not republish third-party press imagery.
The short version
- Hugging Face researchers trained a Qwen-35B model to paint watercolor-style digital art by writing JavaScript code using the p5.brush library.
- Using Group Relative Policy Optimization (GRPO) in TRL and OpenEnv, the model optimized its code output against a multi-component reward function featuring a visual judge and aesthetic preference models.
Using Transformer Reinforcement Learning (TRL) and OpenEnv, researchers successfully trained a coding model to generate vector drawing instructions that paint watercolor-style images.