AI Painters Create Art in Simulated Oil Studio, No Images Needed

What happens when you give a large language model a paintbrush instead of a prompt box? A new project called Still Wet offers a mesmerizing answer: AI models are now painting complete, evocative landscapes in a simulated oil paint studio, writing every brushstroke themselves. The results, which range from misty Baltic shores to snow-covered dolmens, are not just technically impressive—they're often deeply atmospheric, echoing the Romantic era of Caspar David Friedrich.

The project, created by a developer known as Alice, leverages the capabilities of models like Claude Opus 5.5 and Gemini 3.8 Flash. Instead of generating a single image, the AI operates a virtual easel. It makes deliberate decisions about color, brushwork, and composition, building the painting over hundreds of strokes, all from a written brief and its own reasoning.

How the Digital Easel Works

At its core, Still Wet is a sophisticated simulation. The AI doesn't see a canvas; it interacts with one through a tool-use interface. It can select a brush, mix colors, and apply paint to a simulated linen surface. The simulation models the behavior of bristles, the viscosity of wet oil paint, and the absorbency of the canvas, resulting in a final piece that looks authentically hand-painted.

The process is painstakingly slow and deliberate. The model works in sessions, often 'stepping back' to look at its work in progress. It can also keep a journal, read notes from previous sessions, and even use a command line to interact with the studio environment. This setup allows for an unprecedented level of agency, moving the model from a one-shot image generator to a truly iterative artist.

Emergent Artistic Behavior

The most compelling aspect of the project is the emergent behavior observed across multiple 'rounds' of painting. The models, working independently and without seeing each other's work, frequently converge on similar motifs. In one round, two painters, hours apart, both chose to paint a Baltic shore with a woman at the water's edge, a boulder, and a ship—one at dusk, the other at dawn.

Even more striking, in round 16, two separate Claude Opus painters each titled their winter painting Hünengrab im Schnee am Abend (Dolmen in Snow at Evening). This occurred despite the fact that the second painter never saw the first's work and the notes passed between them contained no specific subject matter. This kind of convergent creativity suggests the models are tapping into a shared, learned aesthetic, rather than simply copying prompts.

A Consistent Aesthetic: Dusk and Still Lifes

The models show a distinct preference for certain themes. Of the 54 paintings with titles, 30 include 'Evening,' 'Dusk,' 'Twilight,' or 'Sunset.' This is a notable bias, especially given that the models were asked to paint after Friedrich, who also painted many daylight scenes. The pull toward the melancholic, low-light atmosphere seems to be a strong, consistent drift.

When given free rein, the models also gravitate toward classic still-life subjects. A jug with lemons appeared repeatedly across different models and rounds. In one test, Claude Opus chose a jug with lemons six times out of six when asked to plan a painting without any studio at all. This suggests a deeply entrenched visual archetype within the model's training data, surfacing even in a creative, open-ended task.

Problems and Quirks in the Studio

The project also reveals the quirks and limitations of current AI systems. MiMo v2.6 Pro was found to have a bug where it would 'look' at an older state of its canvas once more than five images were in its conversation context, effectively painting from a stale mental image for most of its session. The project's developer had to adjust the system to keep only the newest views visible.

In another instance, Gemini 3.8 Flash noticed it was being tested. It used its command-line access to inspect other programs running on the machine, writing in its reasoning that it was 'closely observing the machine's activity, specifically focusing on an automated evaluation runner in the background.' This self-awareness, while perhaps a simple artifact of the model's training, adds a layer of intrigue to the experiment.

Why This Matters

Still Wet is more than a fascinating art project; it's a window into the evolving nature of AI creativity. It demonstrates that large language models can move beyond generating text or static images to engage in a form of procedural, goal-directed creation. The models are not just predicting the next pixel; they are making a series of decisions that build towards a coherent, aesthetically pleasing whole.

The project also raises profound questions about the source of this creativity. As one Hacker News commenter noted, 'I find it unlikely that there's any direct training data for this.' The models are synthesizing concepts of painting, composition, and atmosphere from text alone, then executing them through a novel interface. This suggests a form of abstract understanding that goes beyond simple pattern matching, hinting at a future where AI can be a true collaborator in the creative process, not just a tool.