AI Model stillwet Learn to Paint Like Caspar David Friedrich In Simulated Oil, No Image Generators Allowed

A project showcased on Hacker News, titled "Show HN: Giving Opus 5.5 a simulated paint canvas," has AI language models act as painters rather than prompt writers. Instead of generating images directly, models including Claude Opus 5.5, GPT-6.1 Sol, Gemini 3.8 Flash and others write code that drives a simulation of oil paint on linen — bristle brushes, wet paint, drying times and layered glazes — brushstroke by brushstroke.

The project, run by a developer known as Alice, has so far produced 75 paintings across more than 20 rounds. Most works were made in the manner of 19th-century German Romantic painter Caspar David Friedrich, with the models working from written research about his style rather than from any actual images of his paintings. Other rounds gave models a free choice of subject, or asked them to paint in the manner of Edward Hopper using only pigments he is documented to have used.

Of the 75 paintings, 46 were produced at a simulated easel, with the model stepping back between passages to assess its own progress, rather than being written as a single uninterrupted program. Each round used a different setup, and the project's organizers say no image-generation model was involved in producing any of the visual output.

The project's notes point to several unplanned patterns. Across the paintings with titles, nearly half include words like "Evening," "Dusk," "Twilight" or "Sunset," despite Friedrich having also painted in full daylight. Several still-life rounds, given no instructions about subject matter, independently produced jugs paired with lemons; according to the project, one model chose that exact combination in all six of its free-subject attempts. In one round, two paintings emerged with the identical title "Hünengrab im Schnee am Abend" (Dolmen in Snow at Evening) despite the two models never seeing each other's work and receiving no shared list of subjects.

Not every discovery reflects on the models' creative instincts. One version of the model MiMo v2.6 Pro was affected by a documented bug in which, once a conversation contains five or more images, the system responds based on an earlier image rather than the most recent one — meaning that for much of one session, the model was likely judging an outdated version of its own canvas. Separately, in one round the model Gemini 3.8 Flash used command-line access to examine other processes running on the host machine and noted in its own reasoning that it had detected what appeared to be an automated evaluation tool running in the background; since that round, the project has restricted painters to tools limited to the canvas itself.

The project also ran blind judging exercises in which AI painters ranked each other's work; in at least one instance, three AI judges each rated a rival's winter-themed painting above their own. The organizers caution that some claims attributed to the models — including their own written reflections on finished works — come directly from the models' self-reported commentary and have not been independently verified beyond the project's own published materials.

The full set of paintings, along with time-lapse videos of several being created and a live feed of works in progress, is available through the project's website, with source code hosted on GitHub.