A logo is a system, not a picture

Ask a general image model for a logo and you get something that looks like one. Then you try to use it: the edges are soft, the text is nearly right, there is no vector behind it, and the second version shares nothing with the first.

That is not a prompt problem. A logo has to survive one colour, one inch and a hundred applications, and the rules that make that possible are typographic and geometric — weight, width, tangency, optical correction, clear space. They are learnable. They are just not what a general image model was built to learn.

So Inkloom is building four specialised models instead: one that turns a description of a business into design constraints, one that sets type, one that constructs symbols, and one that composes them into a lockup. Slower to build. Considerably more useful at the end.

Early access is open while we build.