Training a coding model to paint watercolours with TRL and OpenEnv

A coding model can be trained to create beautiful, imperfect watercolour paintings by generating JavaScript code. The process uses Reinforcement Learning (RL) to teach the model aesthetic preferences, rewarding it for producing art that aligns with a specific style. This reward system is a sophisticated mix, combining scores from an aesthetic preference model with a vision model that compares new art against a hand-curated pool of reference images. The entire open-source pipeline, from the RL environment to the trained models, is available on Hugging Face to reproduce the experiment.