I've been playing around with generative AI tools for creating articles and other visual aids, but I've had trouble getting Qwen-Image 2512 to define lines and shapes clearly. Here's a process I followed to improve the header image on the Share tokens instead of IDs using the Sam Collins's underdrawing technique.

Underdrawing

The process works by first using a code-to-image technique to build an image (like building an SVG from raw shapes, or a D3 diagram), and then using it as image input for an image-to-image generation model.

Note, in this article I have rasterised all of the SVGs, due to issues that arise with rendering engines and font incompatibility.

Step 1: Generating the base SVG

First, I asked Claude to review my initial header and redraw it as an SVG. This effectively uses its image-to-text and text-to-text capabilities, and it output an SVG matching the base description (a wood-grained table, with a vertically-aligned set of keys on the left, and a pile of emblemed bronze tokens on the right).

Step 2: Revising the SVG

Since SVGs are declarative code, it's easy to iterate on discrete changes without regenerating the whole image (or using inpainting to focus attention).

To save on context, it was beneficial to isolate parts of the SVG, iterate on the details in a separate conversation, and bring them back to the main conversation. Very easy when I can upload and download parts of the SVG as conversation context.

Step 3: Regenerating the drawing

Taking the final SVG, Qwen-Image-Edit in ComfyUI can apply the "shaded drawing" look and feel, without the pressure of needing to define any shapes or object boundaries.

If you have feedback or questions about this article, let's catch up via email.