Using "underdrawings" for accurate text and numbers
AI image generators notoriously struggle with accurate text and numbers, a common headache for users trying to produce precise visual content. This post introduces a clever 'underdrawing' technique that combines deterministic outlines with generative AI to reliably render perfect numbers and text. The simple method surprisingly outperforms even recently released, state-of-the-art models like ChatGPT-Images-2, offering a practical workaround for a persistent AI limitation.
The Lowdown
AI image generation models, despite their impressive capabilities, consistently falter when asked to produce accurate text or sequences of numbers. Sam Collins introduces an ingenious, simple method dubbed the 'underdrawing' technique, which effectively bypasses this limitation by leveraging the strengths of both deterministic and generative systems.
- The core problem is demonstrated with examples where both Gemini 3.0 Pro and ChatGPT-Images-2 fail to correctly render 50 consecutively numbered stepping stones in a spiral.
- The 'Underdrawing Method' posits that deterministic tools excel at precision (like SVG for numbers), while generative AI excels at aesthetics.
- The technique involves two main steps: first, creating a precise 'underdrawing' (e.g., an SVG) with the desired numbers or text in the correct layout. Second, this underdrawing is then passed to an image generation model (like Gemini 3.0 Pro) with a prompt to transform it stylistically.
- This allows the generative model to 'paint' over the accurate outline, resulting in a visually rich image that perfectly preserves the original text and numerical integrity.
- Collins notes the surprising effectiveness of this basic method, as it delivered accurate results where newer, highly capable models alone could not, suggesting a fundamental and accessible workaround for a common AI shortcoming.
This straightforward, two-step process offers a powerful solution for anyone needing reliable text or numerical representations within AI-generated images, demonstrating a smart way to combine the best of different computational approaches.