Using underdrawings for accurate text and numbers in AI art

If you’ve spent any time trying to get an AI model to render a specific sequence of numbers or a complex layout, you know the pain. You ask for a numbered game board, and you get a beautiful, artistic mess of gibberish symbols that look like numbers but fail the moment you try to count them. Most people blame the model, but the truth is that we’re asking generative models to do math—a task they aren't built for.

Here is the reality: generative models are painters, not architects. If you want using underdrawings for accurate text and numbers, you have to stop treating the model like a calculator and start treating it like a digital artist who needs a sketch to trace.

The deterministic advantage

The secret to fixing this isn't a better prompt; it’s a hybrid workflow. You need to separate the layout from the aesthetic. Deterministic tools like SVG, Python, or even a simple Mermaid diagram are perfect for precision. They don't hallucinate coordinates. By generating a simple, high-contrast "underdrawing" of your layout first, you provide the generative model with a structural map.

Once you have that map, the generative model doesn't have to "calculate" where the number 42 goes. It just has to "paint" over the pixels you’ve already placed. This is the part nobody talks about: you aren't asking the model to create; you're asking it to stylize your existing data.

How to implement the underdrawing method

You don't need a complex pipeline to make this work. Here is the two-step process I use to get perfect results every time:

1. Generate the skeleton: Use a tool like Claude or a simple Python script to create an SVG or a basic black-and-white image. This file should contain the exact text, numbers, and shapes you need in their final positions.
2. Apply the style: Feed that underdrawing into a multi-modal model like Gemini 3.0 Pro. Your prompt should explicitly tell the model to transform the provided image into your desired aesthetic—claymation, oil painting, or photorealistic—while maintaining the structure of the underdrawing.

Why does this work so much better than standard prompting? Because you’ve removed the cognitive load of spatial reasoning from the generative process. The model is no longer guessing where the spiral winds; it’s simply following the lines you’ve provided.

The edge case nobody mentions

Here’s where most people get tripped up: the model might still try to "improve" your layout. If your underdrawing is too messy, the model will interpret it as a suggestion rather than a constraint. Keep your underdrawings clean, high-contrast, and strictly defined. If you’re struggling with AI image generation consistency, it’s almost always because your source layer is too ambiguous.

Does this method work every single time? No. You’ll still see occasional artifacts or minor misinterpretations, especially with complex fonts. However, it is significantly more reliable than hoping for a lucky roll of the dice from a text-to-image prompt.

If you’re tired of fighting with broken text in your designs, stop relying on the model's internal logic. Build the structure yourself, then let the model do the heavy lifting on the visuals. Try this workflow on your next project and see how much time you save on manual edits. Read our breakdown of advanced prompt engineering techniques next to see how you can further refine your results.