Surya Narreddi: Training AI to Paint with Code

Primary Source: https://surya.website/rling-qwen-to-paint-with-code
Author: Surya Narreddi · Year: 2026 · Publisher: Surya Narreddi

Summary

An empirical study on applying Reinforcement Learning (RL) directly to open-weight LLMs (Qwen) to master programmatic painting in SVG and code, developing visual intuition without pixel diffusion.

Key Points

• Formulates painting as a discrete sequential code generation task optimized via visual reward functions (CLIP similarity and raster rasterization loss).
• The model autonomously discovers brushstroke composition, geometric economy, and abstract stylistic shorthand.
• Demonstrates that RL on executable visual code yields cleaner, scalable vector art than raster-to-vector tracing.

Mentioned Pages

Entities

• Surya Narreddi

Concepts

• Reinforcement Learning for SVG Painting
• Code as Visual Action Space