Using "underdrawings" for accurate text and numbers
2 days ago
- #text-reliability
- #two-step-technique
- #AI-image-generation
- The author discovered a technique to generate reliable text and numbers in AI-generated images, surpassing current state-of-the-art models like Gemini 3.0 Pro and ChatGPT-Images-2.
- The technique involves a two-step process: creating a deterministic 'underdrawing' (e.g., using SVG) with correct text and numbers, then using generative AI to apply a visual style without altering the text.
- An example prompt for a game board with 50 stepping stones in a spiral was used to demonstrate the failure of Gemini and ChatGPT in accurately rendering numbers, while the author's method succeeded.
- The author suggests tools like Claude can automate the steps, making the technique accessible, but notes it's not perfect every time (e.g., a 'reality check' with an error like 71).