Using an "Underdrawing" for Accurate Text and Numbers

Image generation models excel at visual expressiveness, but have fundamental limitations in getting the exact position and order of text or numbers right. To address this, a technique using underdrawing has been proposed that separates the roles of deterministic tools and generative models.

Workflow

• Deterministic layer: Use SVG/HTML, Python, Mermaid, etc. to first generate an image that precisely places numbers and text in the desired position and orientation. This step secures mathematical placement and precision.
• Generative layer: Feed the generated underdrawing image together with a text prompt into a multimodal image model such as Gemini 3.0 Pro. The model applies a visual style on top while preserving the structure of the underdrawing.

This approach, similar to ControlNet or img2img, guides the generative model's output using a code-based precise outline as a reference. While it doesn't guarantee perfect results every time, it's a useful workflow for compensating for model constraints in image generation tasks where the placement of text and numbers matters.