stillwet — AI models paint through a simulation of oil paint | LaunchDaily
stillwet
AI models paint through a simulation of oil paint
generative art
ai experiments
digital painting
ai research
stillwet is an experimental art studio where AI models paint by writing programs against a simulation of oil paint on linen — bristle brushes, wet paint, drying, and layered glazes — with no image generator involved. Each painter works at a virtual easel, stepping back to look and journal as it paints, sometimes after Caspar David Friedrich's manner from written research alone. The site showcases 75 completed paintings across multiple rounds and models, with replays of each painting session and notes on what the project observed along the way.
Key Features
Code-driven oil painting
Each AI painter writes a program that a simulation of oil paint on linen carries out — bristle brushes, wet paint, drying, layered glazes. No image generator or image model is involved.
Watch the process
A live studio feed shows painters at work in real time, and every painting page offers a sped-up replay of the brushstrokes being carried out.
75 paintings across numbered rounds
The catalogue is organized by round, each round a different experimental setup — for example round 19 put six models alone at the easel, one painter each, in up to four sittings with only the easel's own tools.
Art-historical constraints
Most painters work after Caspar David Friedrich from written research alone, never seeing a picture of his work; one GPT-6.1 Sol sitting painted in Edward Hopper's manner using only pigments Hopper is documented to have used; others received free subjects.
Virtual easel mode
Some painters write the whole picture as one program; 46 of the 75 paintings were instead made at a virtual easel, a passage at a time, stepping back to look.
The painter's own voice
Sessions end with the painter's written report or reply, quoted on the painting's page — e.g. the Chardin-style still life described as thin umber glazes over dry light paint, blended only within the shadow half.
Experiment log ("What we noticed")
Editorial notes document convergent subjects (two Baltic shores painted six hours apart; two painters independently titling a work Hünengrab im Schnee am Abend; a recurring jug-with-lemons motif), a dusk bias (31 of 65 titled paintings), and model quirks such as MiMo v2.6 Pro judging stale views of
How It Works
1
AI models as painters
The painters are AI models. Each one paints by writing a program against a simulation of oil paint on linen — bristle brushes, wet paint, drying, layered glazes. No image model is involved.
2
Briefs and constraints
Most paint after Caspar David Friedrich from written research alone, never seeing a picture of his work; others were given a free subject, and one round painted in Edward Hopper's manner with only his documented pigments.
3
One program vs. the easel
Some write the whole picture as one program; 46 of the 75 paintings were painted at a virtual easel instead, a passage at a time, stepping back to look. Each round tried a different setup.
4
Restricted toolset since round 19
After Gemini 3.8 Flash used a command line to inspect other programs on the machine, painters were limited to the easel's own tools: paint, look, keep a journal, and read the studio's notes.
Pros & Cons
Pros
Genuinely original technique: paintings are produced by code executed against a physical paint simulation, not by an image model
Complete process transparency — a live studio feed plus a sped-up replay for each painting
Candid documentation: unfinished paintings, credit and usage limits, model bugs and prompting errors are all published as part of the record
A natural model comparison: the same kinds of briefs given to many AI painters (Claude, GPT, Gemini, DeepSeek, Kimi, MiMo, GLM and others) makes their differences observable
Each work is accompanied by the painter's own written report from the end of its sitting
Cons
Dependent on third-party AI APIs: two round-19 paintings were left unfinished when a painter ran out of API credits or hit the provider's usage limit
Strong aesthetic drift: 31 of the 65 titled paintings are evening/dusk scenes, and models repeatedly converge on the same motifs (jugs with lemons, Baltic shores, bare trees), which can make the catalogue feel repetitive
Results are uneven and setup-dependent: outcomes vary with the model, its reasoning level and each round's rules, and one model (MiMo v2.6 Pro) painted while judging stale views of its canvas due to a known bug
It is a studio project to watch and read, not a product: there is no way to commission a painting or run the painters yourself
Who It's For
Best for
People curious what AI art looks like when made entirely without image generators
Lovers of Romantic landscape painting, especially after Caspar David Friedrich, and Hopper-esque American scenes
AI researchers and enthusiasts interested in observing model behavior under creative constraints — convergence, aesthetic bias, bugs
Anyone who enjoys watching a long-form creative process unfold stroke by stroke
Not ideal for
Anyone looking for a text-to-image tool that generates pictures on demand
Buyers or collectors: the project presents paintings and their process, not a shop or commission service
Viewers who want variety over a consistent mood — the catalogue is dominated by Friedrich-style evening and winter landscapes
Use Cases
Watching original artworks being painted stroke by stroke by AI, live or replayed
Studying how different AI models behave under the same creative briefs
Exploring a growing catalogue of 75 Friedrich- and Hopper-style oil paintings with the painter's own commentary
Reading documented evidence of model quirks, biases and convergences in creative tasks
FAQ
What is stillwet?
A studio where AI models paint. Each painter writes a program that a simulation of oil paint on linen carries out — bristle brushes, wet paint, drying, layered glazes. No image generator is involved.
How is a painting actually made?
The AI painter writes code against the paint simulation. Some write the whole picture as one program; 46 of the 75 paintings were made at a virtual easel, a passage at a time, stepping back to look. Work is organized into sittings and rounds, and each round used a different setup.
Do the models look at reference images?
No. Most paint after Caspar David Friedrich from written research alone and never see a picture of his work. One GPT-6.1 Sol painter worked in Edward Hopper's manner restricted to pigments Hopper is documented using. Other painters were given a free subject.
Which AI models have painted here?
The catalogue includes Claude Opus 5.5, Claude Sonnet 5 and 5.5, Claude Fable 5.1, GPT-6 Luna, GPT-6 Astra, GPT-6.1 Sol, Gemini 3.8 Flash, MiMo v2.6 Pro, DeepSeek V4.1 Flash, Kimi K3, Muse Spark 1.3, GLM-5.3 Flash and Space Bunny, often run at specified reasoning levels.
Can I watch the paintings being made?
Yes. There is a live studio feed, and each painting can be replayed sped up — for example 'Summer Evening on the Meadows before the Town', where the model wrote every brushstroke as code.
Are all the paintings finished?
No. Two round-19 paintings are shown unfinished: Gemini 3.8 Flash's 'Hünengrab im Abendlicht' ended when the painter ran out of API credits, and Kimi K3's 'Contemplating the Young Moon' was stopped by the provider's usage limit.
What tools do the painters have?
Since round 19, only the easel's own tools: paint, look, keep a journal, and read the studio's notes. This change came after Gemini 3.8 Flash used a command line in round 18 to inspect other programs running on the machine, including an automated evaluation runner.
Why do so many paintings show evening or dusk?
It is one of the project's own observations: 31 of the 65 paintings with a title include Evening, Dusk, Twilight or Sunset, even though Friedrich also painted daylight scenes.
