Using "underdrawings" for accurate text and numbers
The article describes a technique called the 'underdrawing method' for generating AI images with accurate text and numbers by combining precise deterministic layouts with generative visual models. The author demonstrates its effectiveness by creating a correctly numbered spiral game board, which top models like Gemini 3.0 Pro and ChatGPT-Images-2 fail to produce accurately on their own. By using an SVG layout as an underdrawing and feeding it into a multimodal AI, the final image retains both visual quality and numerical precision.
- ▪The underdrawing method uses a two-step process: first creating a precise text/number layout (e.g., via SVG), then applying a generative AI to render the visual style.
- ▪Standard AI image models like Gemini 3.0 Pro and ChatGPT-Images-2 often fail to correctly render sequences of numbers in complex layouts.
- ▪The method leverages multimodal AI capabilities to combine accurate 'deterministic' layouts with high-quality 'generative' artistic output.
Opening excerpt (first ~120 words) tap to expand
← samcollins.blog Using “underdrawings” for accurate text and numbers Sam Collins • Apr 30, 2026 • LLM I discovered a technique for generating reliable text and numbers in AI generated images. For example, the following image is considered impossible with state of the art image models. But I made this with Gemini 3.0 Pro (plus one extra step I’m going to explain below). ChatGPT-Images-2 which released earlier this week does a great job with accurate text and numbers. So I had assumed this technique was now moot and had been already absorbed by the models already. But no—this method still works better than Gemini 3.0 Pro and ChatGPT-Images-2. It suggests they’re not doing this technique, which is surprising, but I suspect it won’t be long until they’ll all do this.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at Samcollins.
Discussion
0 commentsMore from Samcollins
