Fully offline · Open source and free · Data stays on your computer

Annotate, train, deploy,
all in one app.

Import images and start annotating. Label a small batch and you can train, then use your own model to pre-label the rest. Evaluation, export and live testing on site all happen in one offline app.

No internet needed; data stays local the whole time One-click training environment setup with automatic GPU detection Apache-2.0 license
01
Import
Images / folders / existing datasets
02
Annotate
Boxes, polygons, classification, keypoints
03
Train
Live loss / mAP curves
04
Evaluate
Metrics, confusion matrix, errors
05
Pre-label
Label the remaining images with your model
06
Deploy
Export models / live test
Why Wulabel

Your whole toolchain in one window

No more shuttling data between annotation tools, training scripts, evaluation code and export converters. Install it and go; the more you label, the less work it takes.

Fully offline, data stays on your computer

Images, annotations and models stay in your own folders the whole time. Nothing is uploaded, no ports are opened, and no cloud service is required, so it's safe for classified and intranet environments. Helper features that need the internet (such as Xuntong AI) are optional and never used unless you turn them on.

Faster the more you label

Label a small batch to start training, and the trained model helps you pre-label the rest, so you only fix mistakes. The more data you have, the more manual work you save.

Zero-config environment

Install the training environment in one click. It detects your GPU, picks the right PyTorch version, and includes China mirrors for faster downloads. No wrestling with CUDA and dependencies.

All the way to on-site deployment

Export to ONNX / OpenVINO, test live on RTSP cameras, and build inference service packages that pull streams and raise alerts out of the box. Not just annotation: it takes you all the way to production.

Health check

Automatically finds tiny boxes, duplicate boxes, near-duplicate images and class imbalance, and can use a model to flag possibly missed and wrong labels. Fix them in one click and retrain.

Open source and free

Apache-2.0 licensed, free for personal and commercial use. Training uses Apache-2.0 YOLOX and BSD-licensed torchvision models, and the SAM and OWL models used for assisted annotation are also Apache-2.0, so the models you train are safe for commercial use.

Features

From a single image to a working model

Since v0.2.0, the installer includes all four tasks (detection, segmentation, classification, keypoints), AI-assisted annotation, team collaboration and Xuntong AI integration. Works on Windows, macOS and Linux.

Four tasks

Beyond object detection: instance segmentation (polygon outlines), image classification (one key per image) and keypoints (editable templates for human pose and more). Every task supports annotation, training, evaluation and pre-labeling.

Data management

Import images, folders or by drag and drop, with content-based deduplication and automatic orientation; import annotated YOLO / COCO / VOC datasets with automatic format detection; train/val/test split and one-click export; video frame extraction and frame extraction from live or playback streams (RTSP, HTTP-FLV, HLS/m3u8, MP4) (saved as they're captured); image quality check; use a model to pick the images most worth labeling first.

Annotation workspace

Draw, select, drag and resize boxes; switch classes with 1–9, undo and redo, copy the previous image, autosave; review / skip, model pre-labeling (single or batch); polygon, classification and keypoint annotation; keyframe interpolation and automatic forward tracking for video frames.

AI-assisted annotation

SAM smart select: one click gives an outline or a tight box. Text prompt: finds objects by class name automatically. Xuntong AI recognizes by your classes and only picks up what you want. You can even tell the AI "a helmet with the strap undone still counts as no helmet", and it turns that into recognition rules and tests them for you on the spot.

Health check

Rule checks for tiny boxes, duplicate boxes, near-duplicate images and class imbalance; model checks for possibly missed labels, wrong classes, inaccurate boxes and extra boxes. Add as annotation or ignore in one click, or jump to the Annotate page with the problem highlighted.

Live team collaboration

Your computer is the host, and members sign in from a browser to annotate together, with no install needed; works on a LAN or online. Task assignment, per-image locking with live sync, review and rejection, online status, workload statistics (export for per-piece billing) and watermarks against leaks. Without a network, split the work into offline packages and merge them back; datasets can also be saved as version snapshots.

Train

YOLO-family YOLOX (Nano / Tiny / S / M / L, with the official Mosaic and MixUp augmentation and COCO pretrained weights) v0.3.0; Faster R-CNN (MobileNetV3 / ResNet50) with live loss and mAP curves, training logs, stopping midway and training history; single-stage RetinaNet / FCOS / SSDLite, Mask / Keypoint R-CNN, classification models; augmentation (presets + preview); continue training from an existing model; import external ONNX models (YOLOv5/v8/v11, YOLOX, RT-DETR, etc.); training comparison.

Evaluate

mAP50, mAP50-95, precision, recall, per-class metrics, confusion matrix; browse errors and jump straight back to fix the annotation; threshold recommendation; false-positive tests (run a video with no targets, get false alarms per hour, and add false-positive frames as negative samples in one click); re-evaluate with current data; export a delivery report.

Deploy

Live test with RTSP cameras / local videos and add snapshots back to the dataset; export ONNX / OpenVINO and TensorRT / RKNN conversion packages; build inference service packages (multiple streams, alert rules, Webhook / MQTT push, HTTP API, field data loop); upload to the Xuntong AI platform in one click, and fetch platform alert images back for annotation.

Privacy anonymization

When exporting a dataset or splitting into packages, apply mosaic / blur / black-out to faces, license plates and other areas, either by the classes you labeled or by automatic detection. Only the exported copies are changed; the originals stay untouched.

Latest version v0.6.0

Download and get started

Download, install and go. On first launch, it guides you through a one-click training environment install. If you have v0.2.1 or later installed, the app will prompt you to upgrade; from v0.4.1, click "Update now" and it downloads, installs and restarts automatically, with nothing to do by hand.

SystemSupported versionsInstaller
WindowsWindows 10 / 11 (64-bit).exe installer
macOSmacOS 11 or later, universal for Apple silicon and Intel.dmg
Linuxx86_64: Ubuntu 22.04 or later, Debian 12, UOS, Kylin / NeoKylin (x86 computers), etc..deb, .AppImage
Linux ARM64Kylin, UOS and others on ARM computers such as Phytium and Kunpeng (from v0.4.2).deb, .AppImage
Not supported yetLoongson, HarmonyOS. HarmonyOS devices can join team collaboration in a browser—
Training: an NVIDIA GPU or Apple silicon is recommended, but CPU-only also works License Apache-2.0

First launch: if Windows says "Windows protected your PC", click "More info → Run anyway"; on Mac, right-click Wulabel in Applications → Open; on Linux, install the .deb with sudo apt install ./filename.deb, or make the .AppImage executable (chmod +x) and run it directly.

Questions or feature requests? Email the author.

Previous versions

See GitHub Releases for what's new in each version.

FAQ

You might be wondering

Is it really completely free?

Yes. The app is open source under the Apache-2.0 license, free for personal and commercial use, with no feature paywalls, no accounts and no online license checks.

Will my images and data be uploaded?

No. All images, annotations and trained models are stored in project folders on your own computer, and the app doesn't need the internet or open any network ports to run. It only goes online when you choose to install the training environment or download pretrained weights and assisted annotation models, and images are sent to the Xuntong AI platform only when you choose to use "Xuntong AI" assisted annotation. The app also checks for updates once at startup, sending only the version number, OS type and a random anonymous ID (to estimate roughly how many people use it), with no images or personal information. You can turn this off in Settings.

Can it annotate videos?

Yes. Import local videos and extract frames at an interval, or connect directly to an RTSP camera or a live / playback URL (HTTP-FLV, HLS/m3u8, MP4) and save frames as they're captured. When annotating, you can interpolate automatically between two keyframes, or track forward automatically from one frame, instead of drawing frame by frame.

What about annotating as a team?

Live collaboration is supported: your computer is the host, and members open an address in a browser and sign in to annotate, with no install needed. On the same LAN they connect directly; outsourced members on the internet can go through your own relay server or the Xuntong official relay. You can assign tasks, review and reject, and track each person's workload. Without a network, you can also split the work into offline packages and merge them back when done.

Does training use YOLO? Which version?

From v0.3.0 you can train YOLOX from the YOLO family, in five sizes: Nano, Tiny, S, M and L. Nano and Tiny suit CPUs and NPU edge boxes, S offers the best balance of speed and accuracy, and M and L suit setups with a GPU that need higher accuracy. It uses the official YOLOX Mosaic, MixUp and other augmentation plus COCO pretrained weights. Trained models can be evaluated, used for pre-labeling, exported to ONNX / OpenVINO / TensorRT / RKNN, or uploaded straight to the Xuntong AI platform.

YOLOv5 / v8 / v11 aren't used because they're licensed under AGPL-3.0, which restricts use in commercial products; YOLOX is Apache-2.0, so the models you train are safe for commercial use. YOLOv5 / v8 / v11 models trained elsewhere can also be exported to ONNX and imported into Wulabel for pre-labeling and evaluation. You can also train Faster R-CNN, RetinaNet, FCOS, SSDLite, Mask R-CNN and other models.

Can I use it without a GPU?

Yes. Annotation, health checks and export don't need a GPU. Training also runs on a CPU, just more slowly; an NVIDIA GPU is used automatically if you have one.

Which dataset formats are supported?

Import and export both support the three mainstream formats YOLO, COCO and VOC, and the format is detected automatically on import. YOLO-seg / pose, COCO segmentation / keypoints and class folders are supported too. Models can be exported to ONNX, with conversion for OpenVINO, TensorRT and RKNN; you can also import ONNX detection models trained elsewhere (YOLOv5/v8/v11, YOLOX, RT-DETR, etc.) for pre-labeling and evaluation.

How is it different from online annotation platforms?

It puts annotation, training, evaluation, pre-labeling and deployment into one offline app, so you don't have to upload data to the cloud or build a training environment yourself. The model you train on your own data comes right back to help you pre-label, so you get faster the more you use it.

Windows says "Windows protected your PC" or the browser says the file is unsafe. What should I do?

This happens because the installer isn't code-signed yet and hasn't been downloaded by many people, so Windows and browsers don't recognize the publisher. Every newly released small app gets this; it's not a virus. The installer is only offered on this site and GitHub, and all the source code is public.

If the blue "Windows protected your PC" screen appears during install: click More info, then Run anyway. If the browser warns that the file "isn't commonly downloaded" or "could harm your device": click ⋯ → Keep in the download list (in Edge, also click "Keep anyway"). If security software such as 360 or Tencent PC Manager blocks it, choose Allow or add it to the trusted list.

Which systems are supported? Does it work on Linux, Kylin, UOS or HarmonyOS?

Windows 10 / 11 (64-bit) and macOS 11 or later (universal for Apple silicon and Intel). From v0.4.0 there's also a Linux version (x86_64, .deb and .AppImage) that installs on Ubuntu 22.04 or later, Debian 12, UOS, and Kylin / NeoKylin on x86 computers; from v0.4.2 there's also an ARM64 version for Kylin and UOS on Phytium and Kunpeng computers. Loongson and HarmonyOS aren't supported yet, but HarmonyOS devices can join team collaboration in a browser.

Does it support macOS?

Yes. There's been a Mac version since v0.2.0 (macOS 11 or later, universal for Apple silicon and Intel), and training on Apple silicon uses MPS acceleration automatically. If the first launch says the developer cannot be verified, right-click Wulabel in Applications → Open; if it says the app is damaged, run xattr -cr /Applications/Wulabel.app in Terminal and open it again.

Contact

Contact 小物学长

Wulabel is developed and maintained by 小物学长. If you run into problems, want a new feature, or want to talk about deploying AI recognition projects or custom models, feel free to reach out.

Email: 177183719@qq.com (questions, feature suggestions and project collaboration are all welcome)

Email

Email

Questions, feature ideas or collaboration — just send an email. We usually reply within a day or two.

177183719@qq.com
GitHub

GitHub Issues

You can also report problems and suggest features as a GitHub issue, where everyone can follow the progress.

Open an issue on GitHub