The original title is "Code or diffusion? A field guide to programmatic image generation"

Short summary

A practical field guide for deciding whether to generate images via diffusion models or via LLM-written code rendered deterministically. The key test: if you can describe it with coordinates, shapes, and labels, generate code (SVG, draw.io, Canvas, etc.); if it needs photoreal texture, use diffusion. Code-based rendering costs less, is editable, diffable, and enables machine-checkable validation that bitmaps cannot offer.

• •Structure (diagrams, charts) should use code rendering; texture (photos) should use diffusion
• •Code artifacts are repeatable, editable, and testable — bitmaps are not
• •Renderer tier selection from static SVG to Three.js to Godot covers most structured visual needs

Generated with AI, which can make mistakes.