Sergio Paniego published an open training experiment on September 3, 2026, in which a coding model writes JavaScript to paint watercolours. It builds on earlier work by artist Surya Narreddi, whose August demonstration showed a language model creating illustrations through the p5.brush drawing library.
Paniego reproduced the approach with Hugging Face's TRL and OpenEnv tools, publishing the training scripts, model adapters, reference images and generated results. The model produces programs of roughly 150 lines of JavaScript; each program renders a painting that can be inspected or edited.
Instead of asking a model to copy a fixed picture, Paniego manually rated 178 model-generated reference paintings as ones he loved or considered acceptable. The training reward combined visual preference models with a judge comparing each new painting to that collection. In his three runs, average reward increased, and the models produced fewer near-blank images. Those improvements reflect the experiment's chosen scoring system and the author's taste.
The published recipe still requires substantial hardware. Paniego reports about 18 hours on an H200 GPU for a 60-step run, or roughly 34 hours for 110 steps, plus a separate scoring service. This is a reproducible art-training workflow, not a one-click painting app.