A new project demonstrates a different approach to AI-driven art. Rather than using image generation models, the participating AI models "paint" by writing code that manipulates a simulation of oil paint on linen. The simulation mimics real-world physical properties, including bristle brushes, wet paint, drying times, and layered glazes.
AI Models "Paint" Using Code and an Oil Paint Simulation
In experimental rounds, models such as Claude Opus 5.5, Gemini 3.8 Flash, and GPT-6.1 Sol were given tasks, such as painting in the style of Caspar David Friedrich or responding to specific subjects. While some models produced their entire work as a single program, 46 of the 75 paintings were created using a "virtual easel" method, where the model paints the image in stages and steps back to observe the canvas.
The project highlights how different models approach visual composition and subject matter through logic and programmatic instructions. Some models showed tendencies to repeat certain motifs, such as specific still-life arrangements, based on their reasoning and session notes.
Sources
- Show HN: Giving Opus 5.5 a simulated paint canvas （Hacker News Frontpage, 2026-10-02）
- source on GitHub