Using “underdrawings” for accurate text and numbers
Summary
The article introduces an “underdrawing” technique to improve accuracy of text and numbers in AI-generated images by using a deterministic, initial layer (SVG/outline) that places the text precisely, followed by a generative layer to render the final visuals. It provides concrete examples comparing results with and without the underdrawing across models like Gemini 3.0 Pro and ChatGPT Images-2, arguing the two-layer approach yields superior reliability and may become a standard workflow.