Turn your photos into a YOLO dataset

What the YOLO export is

CutLift exports YOLO segmentation labels: each line is a class index followed by the normalised x and y points of one polygon. That is the format YOLO segmentation models train on. Coordinates are divided by the photo’s width and height, so they run from 0 to 1 and stay valid when the image is resized for training. It does not write the box-only detection format. If you need boxes, the COCO JSON export holds a bounding box for every object, and a box can also be read from each polygon’s minimum and maximum points.

Building the folder

Each photo is exported on its own, as a ZIP with labels/photo-name.txt and a classes.txt. The label file takes the photo’s name, so the pair matches without renaming. To build a YOLO dataset from images, collect the label files in one labels folder, put the matching photos in an images folder with the same names, and split both into train and val sets. Your training tool also needs a small dataset file listing those folders and the class names; you write that yourself.

Class numbers that stay the same

Your class list carries over to the next photo, and to your next visit, so “person” is the same number in every label file and every classes.txt matches. The export window shows the ids before you download. Adding to an existing dataset? Choose Load class list in the class bar and pick its classes.txt, data.yaml or COCO JSON: its order comes first, so new labels line up with the old ones.

The class bar keeps names consistent, and number keys 1 to 9 apply a class to the selected objects. Objects without a class are left out of the YOLO file, so nothing lands under the wrong number.

Pro, with a free try

The YOLO dataset maker is a CutLift Pro export, at $7.99 a month or $47.99 a year. Every account gets one free premium download, enough to check that the labels load in your training pipeline before you pay. AI select (Pro), Quick Selection, the lassos and the Pen all produce outlines that export the same way. Mask PNGs stay free. Labelling and export run in your browser, so the photos in your dataset are never uploaded.