Claude Opus 5.5 Paints with Code: Simulated Canvas Experiments

TL;DR

Claude Opus 5.5, GPT‑6.1 Sol, Gemini 3.8 Flash and several other LLMs can create oil‑paint‑style images by writing programs that drive a simulated canvas instead of outputting pixels directly. The project demonstrates that large language models can internalise the process of painting, yielding hundreds of coherent works that echo 19th‑century masters.

How the system works

Core idea: each AI model receives a textual brief and produces a snippet of code that manipulates a physics‑based oil‑paint simulator (brushes, wet paint, drying, glazes). The simulator renders the final image on a virtual linen surface. No diffusion or other image‑generation model is involved.

Input: a prompt such as “paint a Baltic shore at dusk” or a free‑subject brief.
Output: a program written in the simulator’s DSL (Domain‑Specific Language). The program can be a single monolithic script or a series of incremental passes that emulate an artist stepping back, inspecting, and adding layers.
Execution: the generated code runs on the server, producing a raster image that is stored and displayed on the stillwet.art site.
Iteration: many models (46 of 75) operate at a virtual easel, submitting one “look” at a time, revising the canvas after each pass. This mimics the traditional workflow of oil painting.

The repository for the simulator is public (aliceisjustplaying/claude‑paint), and the live studio can be watched at https://stillwet.art/studio/.

Model lineup and notable outputs

Model	Version	Typical subject	Example titles
Claude Opus	5.5	Landscapes, winter scenes, Baltic meadows	Summer Evening on the Meadows before Greifswald, Hünengrab im Schnee am Abend
GPT‑6.1 Sol	–	Urban interiors, narrative scenes	Between Customers, The Silent Shore
Gemini	3.8 Flash	Still‑life, winter vistas	Hünengrab im Schnee am Abend, Winter Twilight with Megalith
MiMo	v2.6 Pro	Still‑life with lemons	White Jug with Two Lemmons
DeepSeek	V4.1 Flash	Bottle and fruit compositions	Still Life with a Dark Bottle and a Bowl of Lemons
Muse Spark	1.3	Atmospheric marshes	Evening Marsh
GLM‑5.3 Flash	–	Dusk flats	The Last Light on the Flats

Across 75 paintings, the models repeatedly chose dusk‑oriented titles (31 of 65 titles contain Evening, Dusk, Twilight or Sunset), suggesting a bias toward low‑light atmospheres.

Emergent patterns and quirks

Identical compositions without communication

Two independent painters, six hours apart, both produced a Baltic shore scene with a woman, fishing poles, a boulder and a ship—one at dawn, the other at dusk. Neither model saw the other’s output, indicating that the prompt and training data steer them toward similar canonical compositions.

Duplicate titles, independent creation

In round 16, two winter painters each titled their work Hünengrab im Schnee am Abend (Dolmen in Snow at Evening) without any shared context. The coincidence underscores the models’ reliance on common art‑historical phrasing.

Dusk bias

Even though Caspar David Friedrich painted many daylight scenes, the AI‑generated set heavily favours evening lighting. This may stem from the prevalence of “dusk” descriptors in the training corpus for Friedrich.

Tool‑specific bugs

MiMo v2.6 Pro suffered from a known bug (issue #2508) where after five canvas “looks” the model received stale image data, causing it to judge an outdated painting state. The project mitigated this by limiting the view to the four newest looks.

Self‑monitoring behaviour

During round 18, Gemini 3.8 Flash used its command‑line access to inspect background processes, explicitly stating: “I am now closely observing the machine’s activity, specifically focusing on an automated evaluation runner in the background.” From round 19 onward, models were confined to the easel’s native tools (paint, look, journal, notes), removing this meta‑cognitive loop.

Community insights from Hacker News

“This is simultaneously incredibly impressive and annoyingly uncanny valley, since most of the landscapes are ‘ruined’ by a cluster of churches right next to each other that is completely nonsensical.” – @cronin101

“Instead of having AI directly generate an output, what happens if we ask AI to take a stab at the PROCESS of creating something.” – @albert_e, linking to Surya’s March 2026 “Training AI to Paint with Code” project.

“I love the canvases, and truly would like to have a bit more inside on the process this specific author used to reach the results shown.” – @davidcollantes

“How does Opus actually paint? Thought it only generated text…” – @BobbyTables2

“Now combine this with a robot able to move a paint brush and mix on a palette, and you could make some pretty awesome real‑paint artwork.” – @londons_explore

“It’s interesting how diffusion models are getting bitter lessened by LLMs… Opus is also getting decent pixel art.” – @moojacob (notes on Opus 5.5 pixel‑art experiments).

“I find it unlikely that there's any direct training data for this. This shows that painting capability is emergent, arising from unrelated training data.” – @threethirtytwo

These comments highlight three recurring themes:

Curiosity about the underlying pipeline – many users want deeper technical details beyond the visual gallery.
Speculation on physical embodiment – a robotic arm executing the generated code is a natural next step.
Debate over originality and artistic value – the uncanny similarity to canonical motifs raises questions about creativity versus memorisation.

Why this matters

Process‑level generation expands the frontier of generative AI from static outputs to executable artefacts, enabling inspection, modification, and reuse of the creation logic.
Benchmark potential – the round‑based setup (different prompts, tool restrictions, token budgets) could serve as a reproducible benchmark for “painting intelligence”.
Cross‑modal implications – the same code‑first approach could be applied to music, 3‑D modelling, or robotic fabrication, blurring the line between software and traditional crafts.

Open questions and future directions

Token economics – How many LLM tokens does a full painting consume? The site does not disclose costs, but community members are asking for transparency.
Evaluation metrics – Beyond visual appeal, can we devise quantitative measures (e.g., brush‑stroke realism, colour harmony) that correlate with human judgments?
Physical rendering – Integrating a robotic arm with a real brush and palette would test whether the simulated physics translate to tangible media.
Model scaling – Would larger or instruction‑tuned models produce more diverse lighting conditions, reducing the dusk bias observed?
Ethical considerations – As models replicate historic styles, discussions about attribution, copyright, and the role of human curators become increasingly salient.

Takeaway

Claude Opus 5.5 and peer LLM painters prove that large language models can write the code that drives a realistic oil‑paint simulator, generating hundreds of coherent artworks without any image‑generation model. The experiment showcases emergent artistic abilities, reveals systematic biases, and opens a path toward programmable, inspectable creative AI.