Axer

Blog/Engineering

Image to code, then a fidelity loop

You give the agent a screenshot or a Figma frame. It builds the page, takes its own screenshot, and fixes the gaps — up to three rounds.

Axer Team · 15 November 2026

Engineering
Image to code

“Make this” only works if the agent can see what “this” is, and then see what it actually built. The fidelity loop is that second look.

The loop

You attach a screenshot, or a Figma frame when a Figma connection is in place. The agent first scaffolds the page: structure, type, layout, the obvious pieces. Then it opens the preview in the built-in browser, takes a screenshot, and compares it to your reference, section by section. Differences go back into the editor. It will do this for at most three rounds, then stop and show you where it landed.

The point is not a perfect pixel match on the first try. The point is a closed loop that does not wait for you to notice that a hero is too tall or a button is the wrong weight.

Images in the project

When the page needs a photograph, the agent can pull a stock image from Unsplash, or from Openverse and Wikimedia when a key is not required. The source is written into the project, so attribution is not an afterthought.

If a generated image is a better fit, the agent can use the image model we have configured — when that model is available in the account. If it is not, it falls back to stock or to a simple SVG instead of inventing a broken asset.

What we are careful about

This is not a claim that every design arrives finished. It is a working path from a still picture to a real page, with a known number of correction passes and a preview the agent can actually see.

Back to the blog