Developer documents an "underdrawings" technique for generating accurate text and numbers in AI images, historically a weakness in multimodal models. Testing with Gemini 3.0 Pro and ChatGPT-Images-2 shows it outperforms both models' native approaches on complex spatial text layouts. Despite rapid improvements, this simple technique remains more reliable than the latest releases.
Models
Using "underdrawings" for accurate text and numbers
A simple "underdrawings" technique outperforms Gemini 3.0 Pro and ChatGPT-Images-2 at rendering accurate text and numbers in AI-generated images.
Monday, May 4, 2026 12:00 PM UTC2 MIN READSOURCE: Hacker NewsBY sys://pipeline
Tags
models