Exploded views are one of the most useful images a small product team can make. A single picture explains what is inside a gadget, how the parts stack, and which component matters. They are also one of the easiest images for an image model to get wrong: parts melt into each other, labels turn into gibberish, and screws float in random directions.

After a lot of failed drafts, I stopped writing these prompts from scratch and started from a fixed structure. Here are the five slots I fill every time.

1. The object, in one plain sentence

"A compact mirrorless camera body" works better than three adjectives and a brand mood. Name the object first and keep it generic. Real logos and fake certifications are a fast way to get an image you cannot use.

2. A part inventory

List the parts you actually want to see, in order: front shell, lens mount, sensor board, battery, rear shell, screws. Six to ten parts is the sweet spot. Without a list, the model invents components, and invented components are exactly what makes a teardown visual look fake.

3. One explosion axis

Pick a direction and say it: "parts separated along a single vertical axis" or "radial layout around the main board". Mixing directions is what produces the floating-screw chaos.

4. Callout rules

Ask for numbered labels, thin leader lines, and a short parts list in one corner. Keep each label to two or three words. Long labels are where text rendering usually breaks down.

5. Poster framing

Background color, aspect ratio, and how much empty space you want around the object. A clean light-grey background at 3:4 reads like a technical poster; a dark gradient reads like an ad.

Reusing the structure

The point of the five slots is that you only change the variables between runs: swap "camera" for "desk lamp" or "router" and the layout logic stays the same. I keep a handful of these as templates. The GPT Image 2 exploded assembly templates follow the same idea, with part lists, callout styles, and layout direction already split out so you only edit the object and labels.

If you want to try the structure with your own product, the free GPT Image 2 image generator lets you paste the prompt, add a reference photo, and iterate on spacing and label length until the breakdown reads at a glance.

One last tip: inspect the first result for part count before anything else. If the model added or dropped parts, fix the inventory line first. Style tweaks can wait.