Computer vision training without the upload step

Most computer vision work is split across two tools. You annotate in one, export a dataset, then move it into a training script or a hosted service. The handoff is where the time goes: reconciling class lists, rewriting export formats, rebuilding an environment, and discovering three days later that half the masks were saved in the wrong convention.

AnyLearning puts both halves in one desktop application, and runs the whole thing on the machine in front of you.

What it actually does

• Labelling. Boxes, polygons, keypoints and whole-image classes. Segment Anything (SAM and SAM2) runs locally, so you can pre-label an object with one click and correct the outline instead of drawing it.
• Training. Eight project types: object detection, instance segmentation, image segmentation, image classification, handpose classification, keypoint detection, Tabular AI and Text AI. You pick a model size and an epoch count. The run happens in a separate process, on your GPU if you have one and on the CPU if you do not.
• Export. Datasets go out as YOLO, COCO, LabelMe or AnyLabeling. Trained models go out as ONNX, with the config the run used, so whatever runs them next does not need AnyLearning installed.

Because the labels and the training share one project database, the class list you draw with is the class list the model trains on. There is no export step between them to get wrong.

What "offline" means here, precisely

Python, PyTorch and the pretrained weights are inside the installer. Nothing is fetched after installation, and a machine with no network can label, train, evaluate and export.

One honest exception: the app makes a single outbound request on startup, a version check against a static JSON file on our website. It sends nothing about you or your data, and it fails silently when there is no network. That is the only connection the application opens.

Your images, annotations and trained weights are read from and written to local disk. There is no bucket, no ingest, no account and no activation key.

When this is the right shape of tool

Local training earns its place when the data cannot be uploaded. Patient scans, factory-floor footage, defence imagery, anything under NDA, anything a data-protection officer has an opinion about. For that work the upload step is not an inconvenience, it is a blocker, and no amount of encryption-at-rest messaging gets you past it.

It also changes the cost shape. Training uses hardware you already own, so a longer run costs time rather than money, and there is no per-seat or per-hour meter running while you experiment.

When it is not

If you need to train on hardware you do not own, or coordinate annotation across a team with review queues and per-annotator assignment, a hosted platform or a self-hosted server tool is a better fit. AnyLearning is a single-user desktop application. It has no collaboration features, no shared project server and no cloud sync, and adding those is not the direction it is going.

Tools like Label Studio and CVAT are built around that team annotation problem and are worth using when that is the problem you have.

Current limits worth knowing

• Single user, one machine. Projects are local files.
• On macOS, training runs on the CPU. Apple's GPU backend is not wired up yet.
• Model selection is deliberately narrow. Each project type ships a small set of variants with sensible defaults rather than an open-ended model zoo.

Getting started

Install AnyLearning, then work through the tutorial for the task you have. Object detection is the usual starting point, and auto labeling is worth reading early because it removes most of the tedium.