Sakhanda Wire
NVDA $153.64 -1.60% MSFT $506.98 -2.13% GOOGL $279.95 -1.33% META $620.09 -1.01% AMZN $256.57 -0.46%
← Back to the news

Open pipeline trains a coding model to paint watercolours

An open implementation explores whether reinforcement learning can teach a language model to generate watercolour-style paintings. The project, described by Sergio Paniego, follows Surya Narreddi’s earlier work, in which a model writes JavaScript using p5.brush, a library that simulates features such as pigment bleeding, paper texture and brush flow. Narreddi’s video had received more than 1.5M views when the article was written.

The implementation uses TRL and OpenEnv and publishes the reference dataset, environment, training scripts and models. The model is restricted to ten p5.brush methods, while a gate rejects sketches that fail to compile, produce no pigment, use direct p5 calls or attempt to manipulate the scorer. Rewards combine basic validity and length checks with HPSv3 and a pairwise judge, Qwen3-VL-30B-A3B-Instruct. The judge compares each candidate with four references.

The reference pool contains 178 model-generated paintings rated by Paniego in two preference tiers. They were produced from hibiscus photographs by GLM-5.2, Kimi-K3, Qwen3-Coder-Next and Qwen3.5-122B-A10B. Three training runs tested different balances between the two model-based preference signals; according to the article, all runs with the pairwise judge learned.

Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.

Originally published by Hugging Face on

Read the original on Hugging Face ↗

The reporting and the image belong to Hugging Face and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news