Training a coding model to paint watercolours (huggingface.co)

Surya Narreddi has demonstrated a unique intersection of AI and art by training a language model to produce watercolor paintings through JavaScript code using the p5.brush library. The project went viral after Narreddi released a video showcasing the model's ability to create loose, handmade-style paintings, garnering over 1.5 million views. The significance lies in its approach to artificial creative expression—challenging prevailing trends in generative AI, which often produce hyper-realistic outputs. This model merges artistic preference with technical execution, allowing users to observe the decision-making process involved in each brushstroke. The project leverages Hugging Face for end-to-end development, defining a reinforcement learning environment where aesthetic preferences guide the model's training. Key technical details include using a pairwise judge that evaluates style against a curated pool of paintings rated based on personal taste, and using specific methods within the p5.brush library to maintain a watercolor aesthetic. The model can learn from a limited set of visual cues, making it an innovative exploration of how AI can mimic and adapt artistic styles. As the project evolves, it will push the boundaries of machine learning applications in creative domains, fostering discussions about the nature of art and human-like creativity in AI.