Using "underdrawings" for accurate text and numbers
A simple two-step technique for generating AI images with accurate text and numbers: create a precise SVG "underdrawing" with correct layout, then have the AI model paint over it to make it beautiful.
Read Original Summary used for search
TLDR
• AI image models fail spectacularly at text/numbers but excel at aesthetics—this technique combines code precision with AI beauty
• Step 1: Generate an accurate layout in SVG/code with correct text/numbers in proper positions
• Step 2: Feed that "underdrawing" image to a multimodal model (Gemini 3.0 Pro) with your aesthetic prompt—it paints over while preserving accuracy
• Successfully generated a 100-step spiral board game with correct numbering—something "impossible" with standard prompting
• Works because you're using deterministic machines for precision and generative machines for creativity, not fighting their nature
In Detail
The author discovered a practical workaround for AI image generation's most notorious weakness: accurate text and numbers. Standard prompting fails even with state-of-the-art models like Gemini 3.0 Pro and ChatGPT Images 2—they produce visually impressive results that fall apart on close inspection. The solution emerged from trying to generate a 100-step adventure board game: use code for what it's good at (precision) and AI for what it's good at (aesthetics).
The technique works in two layers. First, create an "underdrawing" using deterministic tools like SVG, Python, or Mermaid—layout all text and numbers in their correct positions and orientations, then export as an image. Second, pass this underdrawing image to a multimodal model (one that accepts image+text input and outputs images) along with your aesthetic prompt. The model essentially "paints over" the underdrawing, preserving the accurate text/number positions while adding visual polish.
The results are dramatic: the same prompt that produces gibberish numbers with standard generation produces a perfectly numbered spiral board game with this method. The author notes it's not 100% perfect (showing an example where "71" got mangled), but it's good enough to unlock an entire category of previously impossible generations—board games, infographics, diagrams, anything requiring precise text layout. The implementation is straightforward enough that Claude or Codex can automate the entire process.