The job is not the image
Why AI-generated logos, app icons and restored photos need production-ready formats, sizes and resolution after generation.
- Published
Ask a current generative model for a company logo and it will give you something plausible in a few seconds. Then try to use it.
You need it as vector, because it is going on a sign and on a favicon and on a business card, and a 1024-pixel PNG is none of those things. You need it to sit correctly on a dark background, which means you need an actual transparent edge rather than a matte the model quietly baked in. You need it at 16 pixels, where every interior detail the model was so pleased with collapses into mud.
None of that is a sampling problem. All of it is the job.
The size ladder is not a resize
An application icon is not one image. On iOS it is a set. On Android it is a set with a mask applied to it. On the desktop it is a different set again, and if you are packaging an Electron app you need the platform-specific container formats on top of that. A single source mark has to survive all of it.
Resizing does not survive it. A mark that reads at 512 pixels frequently does not read at 32, because legibility at small sizes is a design decision — strokes get heavier, counters get bigger, detail gets removed — not an interpolation setting. Doing this correctly means deciding what to sacrifice at each step down the ladder, and doing it consistently enough that the result still looks like one mark.
Resolution has a floor, and the floor is physical
Photo restoration has the same shape of problem pointed the other way. A restored photograph destined for a screen and one destined for a 10-inch print are not the same deliverable, and the difference is not a slider. If the output does not clear the resolution the medium requires, the restoration failed regardless of how good it looks in a browser at 50% zoom.
What this implies about how you build
If you take the position that the deliverable is the product, a few things follow that are inconvenient if you were hoping to ship a prompt box.
The system has to know what it is producing before it produces it. “A logo” and “a poster” are not two prompts against one pipeline; they are two different sets of constraints on output format, dimensions, colour handling, and what counts as done.
Generation has to be one step in a longer job, not the whole job. Which means the work has to be resumable, inspectable, and correctable at each stage — you have to be able to fix the third step without regenerating the first.
And the model has to be replaceable. Models improve on a timescale of months. Any architecture that treats a specific model as load-bearing is building on a part that is guaranteed to be swapped out. The durable engineering is on either side of it.
That is the part we are building.
Zero Agentic’s product is in development. Nothing described here is available yet.