Show HN: Giving Opus 5.5 a simulated paint canvas
Claude Opus 5.5 and other multimodal models create oil paintings by writing code against a simulated paint engine, observing their canvas iteratively without using diffusion image generators.
summary is enough The site presents a striking visual showcase of language models painting through code simulation, and the comment thread adds valuable context on multimodal tool use.
summary
A project called Still Wet uses AI models like Claude Opus 5.5 to create paintings by writing procedural code and interacting with a simulated oil paint and linen environment. Instead of relying on traditional diffusion models, the models operate an easel, make brushstroke decisions, and review their progress iteratively using vision capabilities. The resulting landscapes and still lifes demonstrate unexpected spatial and artistic reasoning through tool use. The accompanying discussion explores whether this represents emergent general intelligence or specific reinforcement learning.
key points
- Models paint by outputting code and commands to a simulated oil-paint engine rather than using standard text-to-image diffusion models.
- Painters inspect their work iteratively using a vision look tool, adjusting brushstrokes, glazes, and textures over multiple passes.
- When given free subjects or prompts based on artists like Caspar David Friedrich, models independently gravitated toward repeating specific motifs like evening light and jugs with lemons.
who should read it
AI engineers, developers, and digital artists interested in multimodal model capabilities and creative tool use.
discussion positive
consensus
Commenters are heavily impressed by models performing complex multi-step creative tasks via code execution and vision feedback loops rather than diffusion generation.
pushback
Some express sadness or skepticism regarding the rise of AI art and its potential impact on human artists.
I find it unlikely that there's any direct training data for this. I'm talking about direct training data for using brush strokes like humans to construct an image.
Correct me if I'm wrong but this shows that painting capability is emergent, arising from unrelated training data thus very very compelling evidence that LLMs are actually intelligent.
Argues that painting via programmatic brushstrokes is an emergent capability from unrelated training data, pointing to genuine model intelligence.
Generated by an LLM from the article and thread. Check the sources before relying on it.