Hugging Face researcher Sérgio Paniego has published a fully open-source reproduction of Surya Narreddi's viral watercolor-painting AI project, which showcased a language model generating paintings through JavaScript code. The project achieved 1.5 million views by leveraging the contrast between AI's typical statistical averageness and deliberately loose, handmade watercolor aesthetics. The model writes code using the p5.brush library, with artistic decisions encoded directly in the generated JavaScript.
The core innovation lies in the training methodology: reinforcement learning optimized for aesthetic preference rather than verifiable metrics. Paniego's implementation combines multiple reward signals, including a vision-based pairwise judge comparing candidate paintings against curated reference images, the HPSv3 aesthetic preference model, and structural rewards for code quality. The entire training pipeline runs on Hugging Face infrastructure with all models, datasets, and code publicly available for reproduction.
The project represents a significant milestone for creative AI applications—demonstrating that language models can be effectively trained toward subjective artistic goals rather than purely functional ones. The approach echoes earlier generative art movements like DeepDream and GAN-based experiments, positioning contemporary AI research at the intersection of creative exploration and technical innovation. The open-source release enables researchers and artists to reproduce and build upon the methodology.
Key Points
Language models trained to generate watercolor paintings by writing JavaScript code, achieving loose, handmade aesthetics that contrast with typical AI image generation
Reinforcement learning optimized for aesthetic preference instead of verifiable metrics—proving RL over subjective taste is viable for creative tasks
Reward system combines pairwise visual judging against reference images, HPSv3 aesthetic model, and code quality metrics across multiple training runs
Fully open-source implementation published on Hugging Face with training scripts, datasets, models, and infrastructure available for community reproduction
Represents creative AI application similar to early generative art experiments, trained on curated datasets rather than prompt engineering alone