1Q onecue.ai

About onecue.ai

Who builds this, why it exists, and how it works.

onecue.ai is a set of free tools for getting text out of images: reading it, translating it in place, pulling one field out of a stack of documents, and — a neighbour to the same problem — compositing a product into a photo. It is built and operated by Growdle Inc., a company in Seoul, South Korea.

It started as a travel problem

The specific moment was standing in front of a menu in Japan, holding a phone, and getting nothing useful back. General translation apps are built for typed text or clean signage; a laminated menu photographed at an angle under warm restaurant lighting is a different problem. The text came back partly untranslated, partly wrong, and with the layout destroyed — which matters, because on a menu the layout is half the information: what belongs to which section, which price goes with which dish.

The same thing kept happening with other travel documents. Museum caption cards in vertical Japanese. A product card explaining what a souvenir stone was supposed to mean. Station signage with a mix of Chinese and English. In each case the failure was not that translation is hard — it is that reading a photograph is hard, and everything after it inherits the damage.

So we started at the reading stage instead of the translating one, and the tools grew from there. Image To Text came first, because knowing exactly what the model did and did not read is the prerequisite for trusting anything downstream. Image Translator followed, with the constraint that the result must keep the original layout. Field Extractor came from the same machinery pointed at a different task: many documents, one value each.

How it is actually run

We think the technical choices are worth stating plainly, because they determine what we can and cannot promise:

What we optimise for

One principle decides most of the arguments: an obvious gap beats a confident mistake.

OCR on real photographs is uncertain, and a tool can respond to uncertainty in two ways. It can lower its bar and output something for every region — which produces plausible-looking text that is wrong, and a user has no way to tell. Or it can leave the uncertain parts visibly unread, which looks worse and is easier to catch. We choose the second, report coverage so you can see how much was left out, mark low-confidence lines as unsure, and let you point at anything the model skipped so it can be read directly.

It makes the output look less impressive in a screenshot. It makes it safer to rely on, which is the point when the text you are reading is a menu you are about to order from.

The company

Growdle Inc. builds AI products from Seoul; the company's other work includes pick2sell, an AI-assisted commerce solution, and byLENA, an AI model generation platform. onecue.ai is where the image and OCR work of that group is made available directly, as free tools.

Questions, bug reports and feature requests are all welcome — see Contact for what to include and how quickly we reply.