Still Wet: Painting with AI Reasoning and Code

The Still Wet project features AI models that 'paint' by writing code to simulate real oil painting techniques instead of using traditional image generators. Through 75 works, the project reveals that different AI models often independently choose similar themes, such as evening landscapes and still lifes with lemons. These findings suggest that AI models share common 'artistic' instincts or biases derived from their training on human culture and research.

Key Points

• AI models create art by writing procedural code for a physics-based oil paint simulation rather than using diffusion-based image generation.
• Models demonstrate emergent biases, such as a recurring preference for painting still lifes of jugs and lemons or landscapes at dusk.
• Different AI models often converge on nearly identical compositions and titles independently, despite having no contact with each other's work.
• The project transitions from single-program execution to a 'virtual easel' setup where models paint in multiple sittings and 'step back' to evaluate their progress.
• High-reasoning models can successfully evoke specific historical art styles (like Friedrich or Hopper) using only written descriptions and technical constraints.

Sentiment

The overall sentiment is cautiously enthusiastic. Hacker News largely agrees that the project is technically impressive and worth exploring, while disagreement centers on how much artistic meaning, novelty, and labor value should be attributed to the results.

In Agreement

• The project is compelling because it asks models to execute a creative process through code, tools, and observation instead of producing a finished image directly.
• The iterative canvas feedback loop makes the outputs feel more interesting than ordinary generated images and helps explain why the work can improve over time.
• The experiment demonstrates meaningful tool-use generalization, especially when models translate visual goals into procedural brush actions.
• The system could become useful for artists as a way to explore composition, technique, or visual ideas before committing physical materials.
• Future versions connected to plotters, robotic arms, or real paint could produce compelling physical artwork and educational projects.

Opposed

• Some outputs remain uncanny or visually nonsensical, with repeated motifs that look like model artifacts rather than intentional artistic choices.
• The capability may be less surprising if models have been trained or reinforced on drawing, painting, or visual-tool environments.
• AI-generated process may not satisfy communities that reject generative AI, because the procedure still lacks human artistic intention.
• The project intensifies concerns that commercial art labor will be automated further, harming artists who already face weak compensation.
• A few commenters see the premise as culturally or ethically alienating because it replaces human-to-human artistic expression with machine production.