For graphics, LLMs work better writing code than making pixels

In a practical guide to AI image generation, commenters shared a shift in approach. One said procedural graphics is coming into vogue as LLM tool use improves. Another asks Opus to write Python that creates logos, which they find more consistent and easier to control than generative image models. A third complained that local image models like Flux or Qwen run 1 to 2 years behind ChatGPT's quality and take a long time.

Separately, a Tiling Noise Show HN showed procedural GPU patterns, and a HN poster noted that LLMs plus code is now a viable route for pixel art.

The pattern: when you need repeatable, editable output, have the model produce a program instead of the artifact. You can diff it, tweak it and rerun it.