YOLO-World
Describe it in words (“white horse”) and map each phrase to your classes. Or bring any Ultralytics model of your own.
Downloads its model once (asks first). Runs on a CPU.
Soft to use. Sharp on every object.
Jelibox is a private, local workspace for building object detection and segmentation datasets. Import what you have, label it with four kinds of AI help, see what it really contains, and export it augmented, without your images ever leaving your machine. On your desktop, or from any browser on your network.
ImportAny folder of images with Pascal VOC, YOLO or COCO labels
Auto-labelFour assistants, from zero-shot prompts to SAM 2
AnalyzeA dashboard that shows what your dataset really holds
ExportYOLO, COCO or VOC, augmented, to any drive
Auto-label
Four ways to let a model do the boring part.
Pick a mode when Jelibox starts and change it any time. Press G for the current image, or label the whole folder in one run.
LocateAnything
NVIDIA LocateAnything-3B labels a whole folder from a list of categories, in English or Chinese.
About 7 GB once. GPU with 8 GB+ recommended.
Custom Model
Find people or bottles, then a small model you train gives each one your class: sleeping / working, Coca-Cola / Fanta.
Train once on images you already labeled.
SAM 2 Dynamic
No training. Annotate a few images and SAM 2 learns from each one, then finds the same objects in the next, as boxes or polygons.
One SAM 2 model, 75–224 MB, downloaded once.
Dataset tools
Everything around the labels, built in.
Most annotation tools stop at the box. Jelibox also helps you bring data in, understand it, and ship it.
Import
Bring any dataset in, safely.
Point Add Workspace at a folder. Jelibox searches every subfolder and shows what it found before it copies a single file.
- Images with Pascal VOC, YOLO or COCO labels, or images alone
- Classes are picked up from the annotations, in the order you want
- An optional prefix renames images and labels in order
- Annotations of a class your workspace does not have are flagged first and skipped, never mixed in silently
Analyze
Know your dataset before you train on it.
Select a workspace and open the dashboard: objects per class, size histograms, image resolutions, all in interactive charts.
- A warning appears when objects are smaller than a limit you choose (32 px by default)
- Simulate a model input size such as 640 px and see how small your smallest objects become. It is calculated, so no image is copied or resized
- Compare 320 to 1280 in one table to pick the input size that fits
- Remove the tiny objects in one click, after a clear warning
Export
Export augmented, and keep it where you want.
YOLO, Pascal VOC or COCO with a train / valid / test split you set with a slider.
- A visual augmentation picker: every effect is a tile with a preview from your own pictures. Boxes and polygons move with the picture
- Preprocessing: Auto-Orient and Resize (stretch, fit or letterbox), and the same picker in the Train window
- Every effect has its own strength and chance; you choose how many outputs each training image becomes; a seed makes it repeatable
- Save to any folder, such as an HDD, so archives stay off your SSD. Each export becomes a new
workspace-v5,-v6… folder and nothing is overwritten
Web version
Install once. Annotate from any browser.
Put Jelibox on the strong computer with the GPU and open it from your laptop. Same workflow, same files, nothing leaves your network.
Browser
The whole workflow, on your network.
Open http://localhost:8420 on the server, or http://<server address>:8420 from another device. Sign in, and everything the desktop does is there.
- Sign in with your own account; the admin adds the others
- Annotate, auto-label with all four modes, train, analyze and export, all in the page
- Upload a dataset from your laptop, or pick a folder that is already on the server
- Exports are downloaded straight to your device; training goes on if you close the tab
- Starts by itself: when you sign in to Windows (no administrator rights), or at boot on Linux (at most one sudo, to allow starting at boot)
Server resources
See what your server is doing.
The workspace page shows the computer that runs Jelibox, refreshed every two seconds: CPU use, RAM use, CPU temperature, GPU use, GPU temperature and storage, on Windows and Ubuntu. A value a computer cannot report simply shows n/a.
- Check that nothing else is using the GPU before you start a training
- Watch temperatures during a long run
- See how much room is left on the drive that holds your datasets
The web page uses plain HTTP, so keep it on a network you trust. Choose App + web when the installer asks; App only installs just the desktop window.
Features
Less friction between image and insight.
Everything you need to annotate, inspect, train, and export without sending your dataset through a cloud service.
Local by design
Your images, labels, and models stay on the machine you control. No account, no upload.
Box + polygon tools
Edit coordinates with handles, zoom, visibility controls and class management. Undo and redo everything with Ctrl+Z / Ctrl+Y.
Train in context
Go from annotations to Ultralytics training without leaving your workspace, and without duplicating your source images.
ZeroFill masking
Black out sensitive regions of an image locally before the dataset leaves your hands.
Repeat and auto-label
Repeat the previous image’s annotations, or label the current image, or the whole dataset, with one key.
Light and dark themes
A calm cream interface that is easy on the eyes, with a dark mode when you need it. Fits a 12-inch laptop screen.
Workflow
A loop that gets faster every round.
Label a few images by hand, let a model help with the rest, check the numbers, and keep correcting. Each pass makes the next one quicker.
Import
Add a folder of images, with or without existing labels, from any common format.
Annotate
Draw boxes or polygons on a handful of images with keyboard-first controls.
Assist
Press G and let YOLO-World, LocateAnything, your own model or SAM 2 label the rest.
Analyze
Check class balance and object sizes, and choose a model input size that suits them.
Export
Ship an augmented dataset in the format your pipeline expects, or train right here.
Start locally
Install once. Keep working.
One command, no git needed. It downloads the newest release and sets everything up. Run it again later to update; your datasets and models are never touched.
Windows
Open PowerShell and paste:
irm https://raw.githubusercontent.com/Jelibox/Jelibox-client/main/install.ps1 | iex
Jelibox gets its own Python (via uv) inside its folder; yours is never touched. Windows asks for administrator permission only if the Visual C++ runtime is missing.
Linux
Open a terminal and paste:
curl -fsSL https://raw.githubusercontent.com/Jelibox/Jelibox-client/main/install.sh | bash
Jelibox gets its own Python (via uv) inside its folder; yours is never touched. It asks for your sudo password only if the OpenCV system libraries (libGL, GLib) are missing.
The installer asks what to install: App + web (the desktop app plus the web version, started for you when you sign in or boot) or App only. It asks again at every update, so you can switch.
The installer asks whether this computer has a discrete NVIDIA graphics card and installs the CUDA or the smaller CPU-only PyTorch accordingly.
Prefer to read it first? install.ps1 and install.sh are short and open source.
Install manually with git instead
Linux
- Install git (if you haven't installed git)
sudo apt install git
- Clone this repository
git clone https://github.com/Jelibox/Jelibox-client.git
- Add access to linux installation script
chmod +x jelibox_linux_installation.bash
- Run Jelibox installation script
./jelibox_linux_installation.bash
The installer needs sudo only if the OpenCV system libraries are missing. It installs uv if needed, downloads Jelibox's private Python, creates the jelibox/ virtual environment, installs dependencies, and creates a desktop launcher.
Windows
- Double-click
jelibox_windows_installation.bat. Do not run it as administrator: it asks Windows for permission itself, only to install the Visual C++ runtime, and only when it is missing. - Follow the installer prompts.
- Open the generated
Jelibox.lnkshortcut.
The installer detects an NVIDIA GPU only to suggest an answer, then asks whether this computer has a discrete NVIDIA GPU and installs the matching PyTorch build. GPU detection does not guarantee that the installed PyTorch package has CUDA enabled.