Code as Visual Action Space

Definition

The paradigm in which generative AI systems interact with visual art not by predicting pixels, but by authoring and executing declarative programs that render imagery.

Key Characteristics

• Preserves topological structure, semantic grouping, and parametric modifiability.
• Leverages compiler feedback and visual execution environments.

Applications

• Procedural graphics
• Computational CAD
• Generative vector art

Mentions in Source

• “Demonstrates that RL on executable visual code yields cleaner, scalable vector art than raster-to-vector tracing.” — surya-narreddi-training-ai-to-paint-with-code