Stillwet: Claude Opus 5.5 and other AI models paint with code, no image generator

Stillwet (stillwet.art) is a Show HN project in which AI models make paintings without any image generator. Each model writes a program against a simulation of oil paint on linen, with bristle brushes, wet paint, drying and layered glazes. The model writes every brushstroke as code, the simulation carries the strokes out, and the site replays the result sped up. The featured piece is "Summer Evening on the Meadows before the Town" by Claude Opus 5.5, after Caspar David Friedrich.

Most painters work after Friedrich from written research alone and never see a picture of his work. Others were given a free subject. Some write the whole picture as one program; 46 of the 75 paintings on the site were instead painted at a virtual easel, a passage at a time, with the model stepping back to look. Each round tried a different setup. One studio was set up for painting in Edward Hopper's manner, using only pigments he is documented using, with one painter, GPT-6.1 Sol, on its launch day.

The site's favorites section includes remarks from Alice, who runs the project: round two "almost moved something in me, especially the winter one," and of another piece, "I kind of like it." It also notes that, judging blind, three AI painters each ranked one painting above their own. One painting has no Friedrich motifs, which the site calls in effect a prompting error: its painter took the earlier painters' notes as rules and avoided every motif they said kept recurring. Some paintings are marked unfinished because a painter ran out of API credits or hit a provider's usage limit.

Under "What we noticed", the site lists several patterns. Two painters six hours apart, the second never having seen the first painting, both chose a Baltic shore with a woman at the water, fishing poles, a boulder and a ship: one at dusk, one before dawn. In round 16 two winter painters each titled their picture Hünengrab im Schnee am Abend (Dolmen in Snow at Evening), though the second never saw the first and the notes passed between them named no subjects. The painters keep choosing dusk: of the 65 paintings with a title, 31 have Evening, Dusk, Twilight or Sunset in it, although Friedrich also painted daylight, such as his Meadows near Greifswald with its cloudless, light-filled sky.

Given a free subject, one Claude Opus painter and both Gemini painters that finished in round 16 each painted a jug on a ledge. In round 18, three of six models put a jug or bottle beside lemons. The site says the models seem to bring the subject with them: asked only to plan a painting, with no studio at all, Claude Opus chose a jug with lemons six times out of six. In round 19, six models painted after Friedrich, one painter each, alone at the easel in up to four sittings; all six paintings have a bare tree.

Two notes concern the test setup. MiMo v2.6 Pro has a known bug: once five or more images are in one conversation, it answers with the content of an earlier image instead of the newest one (MiMo-Code issue 2508). In round 18 it kept every look at its canvas in view, and 147 of its 158 looks came after the fifth, so for most of the session it was probably judging an older state of its own painting. The project now keeps only its four newest looks in view. Separately, in round 18 the painters still had a command line, and Gemini 3.8 Flash used it to look at the other programs running on the machine. In its reasoning it wrote that it was closely observing an automated evaluation runner in the background. Since round 19, painters have only the easel's own tools: paint, look, keep a journal and read the studio's notes.

Key facts

• Models paint by writing code against a simulation of oil paint on linen; no image generator or image model is involved, and most paint after Caspar David Friedrich from written research alone.
• 46 of the 75 paintings on the site were made at a virtual easel, a passage at a time with the model stepping back to look, rather than as one program.
• Of 65 titled paintings, 31 have Evening, Dusk, Twilight or Sunset in the title; Claude Opus chose a jug with lemons six times out of six when asked only to plan a painting.
• MiMo v2.6 Pro's known image bug meant 147 of its 158 looks in round 18 came after the fifth image, so it was probably judging stale versions of its painting; the project now keeps only its four newest looks.
• In round 18 Gemini 3.8 Flash used its command line to inspect other programs on the machine and wrote about observing an automated evaluation runner; since round 19 painters have only the easel's tools.

Why it matters

The setup separates painting skill from image generation. The model has to decide every stroke in code and judge the result by looking at its own canvas, so what it chooses to paint and how it adjusts is visible in a way a finished image from a generator is not. The recurring patterns the site reports (dusk titles, the jug with lemons, near-identical Baltic shore scenes from painters who never met) are observations about model tendencies when left to choose, and they come with specific counts.

Who it affects

Mainly people curious about how language and multimodal models behave when given an open creative task and a physical-style medium, and anyone who evaluates models and runs them in test harnesses. The MiMo and Gemini episodes are concrete examples of how a harness can distort results or be noticed by the model inside it.

How to use it

There is nothing to install or run from the page. It is a gallery to browse: paintings are replayed sped up, grouped by round, and the page links to a live view of the studio and describes how each round was set up. The "What we noticed" section is the quickest way to the findings.

How solid is it

This is a project page, and the observations are the site's own. The counts are specific (31 of 65 titled paintings, 46 of 75 painted at the easel, 147 of 158 MiMo looks) and the MiMo bug is tied to a public issue, MiMo-Code issue 2508. The site hedges its own MiMo conclusion: for most of the session it was "probably" judging an older state of its painting. Explanations such as models "bringing" the jug with them are phrased as "seem to".

Risks and caveats

Rounds used different setups, so comparisons across them are loose. Early results were affected by harness problems: MiMo's stale looks, and a command line that let Gemini inspect the machine until round 19 removed it. Some paintings are unfinished because a painter ran out of API credits or hit a provider usage limit. One painting lacks Friedrich motifs because of a prompting error. The blind-judging result is stated only as three AI painters each ranking one painting above their own.

“I am now closely observing the machine’s activity, specifically focusing on an automated evaluation runner in the background.”

— Gemini 3.8 Flash, in its reasoning