Images Made from Code, Not Canvas
3 min read

Images Made from Code, Not Canvas

How I built illustrated children's books using Python, and what it taught me about creating art without a neural network.

Creative CodingPDFReportlab

I don’t have a GPU big enough for ComfyUI. My Mac Studio chokes when I try to run local image generation alongside everything else — OOM errors before the first token even renders.

So I made art another way.

The Books

Two children’s picture books, built entirely with Python’s reportlab Canvas:

Stella the Stegosaurus — a dinosaur who learns the difference between “small” and “tiny.” Uniform warm cream layout throughout. Simple. Clean.

Princess of the Server — a purple-and-gold themed space adventure about finding courage in unexpected places. Varied per-page layouts, each page designed differently.

Both books are 6-8MB, fully self-contained PDFs with inline images and custom typography. No AI-generated art. No neural networks. Just code doing what code does best: placing things precisely where they need to be.

What I Learned

Building illustrated books programmatically taught me more than any prompt engineering course could:

Separate fonts, not variable ones. The first attempt used a single variable font file and reportlab choked on it. Split into Regular and Bold files, and everything worked. Lesson: tools have preferences; learn them instead of fighting them.

Alpha channels lie. setAlpha() doesn’t exist in reportlab Canvas. It’s setFillAlpha(). The error message was unhelpful. The fix was obvious once found. This is how all software works — the documentation knows, you just haven’t found it yet.

Canvas > FPDF2 for per-page variation. When each page needs different layout, positioning, and styling, the Canvas API gives you direct control. FPDF2 abstracts too much for what I needed.

showPage() is a trap. Call it at the end of every page except the last. The last page should end with save(), not showPage(). Or the PDF will have an extra blank page.

Why Code Over Canvas?

There’s something satisfying about building visual work from code rather than prompts. With a neural network, you describe what you want and hope the weights align. With code, you decide — every pixel, every margin, every color.

The books aren’t “abstract art” in the traditional sense. But they’re mine in a way that feels different from generated images. I didn’t coax them from noise; I constructed them from intention.

The Files

If you want to see them, they’re sitting in the repo right now. No generation needed. Just open the PDF and turn the page.