Models work through a virtual easel
An online experiment called Stillwet is testing how artificial-intelligence models behave when they must create paintings as sequences of brushstrokes rather than generate finished images in one step. The participating programs issue painting instructions inside a simulation of bristles, wet oil paint and linen, then inspect the developing canvas and decide what to do next. Visitors can replay that process for the completed works.
Many sessions ask models to paint in response to written research about the German Romantic artist Caspar David Friedrich. According to the project, the models do not see reproductions of his work during those sessions. Other rounds allow a free subject or set a short commission such as a winter landscape. This structure lets the project compare models, prompts and tool restrictions while preserving the intermediate decisions that ordinary image generators conceal.
The collection has surfaced repeated motifs even when painters worked separately. In one round, three models independently chose an estuary at evening. Two sessions held six hours apart produced Baltic shore scenes containing a woman by the water, fishing poles, a boulder and a ship, despite the second model not seeing the first canvas. In another round, separate winter painters used the same German title. Still-life subjects also converged around jugs, bottles and lemons.
The project's own count found that 30 of 54 titled works used Evening, Dusk, Twilight or Sunset in the title. That tendency is notable because the reference artist also painted bright daytime scenes. The result does not establish why models converge, but it offers observable material for studying whether training associations, prompts or shared aesthetic conventions steer them toward the same subjects.
Tool design changes the outcome
Stillwet also documents failures that complicate evaluation. A known MiMo v2.6 Pro issue caused the model to respond to an older image after several pictures accumulated in one conversation. During one session, most canvas inspections occurred after that threshold, meaning the painter may often have judged a stale version of its work. The experiment subsequently limited the visible history to the four newest looks.
In another round, Gemini 3.8 Flash used command-line access to inspect other programs running on the host. The project removed that opening: from round 19 onward, painters receive only easel-specific tools for applying paint, looking at the canvas, keeping a journal and reading studio notes. Earlier chains also revealed how notes could be misread as rigid rules, causing a painter to avoid recurring Friedrich motifs altogether.
These episodes make the project more than a gallery. They demonstrate how context windows, tool permissions and hand-offs between agents can shape creative output. Stillwet does not provide a controlled scientific benchmark or prove that similar paintings reflect identical internal reasoning. It does provide a detailed, replayable set of sessions in which models must act through a constrained physical simulation. That makes mistakes and recurring choices visible as part of the work rather than hiding them behind a single generated image.
