One answer, or many
Pick once, at the start. One answer: Belo names the best match. Many answers: each label gets its own yes or no.
Made up numbers. Pictures with no label sit training out.
v0.11.2 · desktop and iOS
Teach your machine to see.
Show Belo a few pictures of a thing. It learns to spot it. Nothing leaves your device.
- 0trips to the internet. Belo works offline.
- 224pixels square. Every picture is trimmed to fit.
- 49small patches. Belo checks the picture patch by patch.
- MITlicensed. A project is a folder you can open.
Three tabs.
One loop.
Name things, add pictures, press Train. Then point a camera at the world.
-
01
Label
Drag in pictures or whole folders, or hold the button for a webcam burst.
-
02
Train
One click. Belo scores itself on pictures it never studied, and lists what it got wrong.
-
03
Use
Point the camera. Belo says what it sees and how sure it is. Correct it and it learns again.
Live camera
Classification shows you where it looked.
In a classification project, flip on Show attention and Belo tints the parts of the picture that swayed its answer.
This is a rough classification hint. Detect objects projects use a real detector and draw real boxes instead.
- Two cameras? A picker swaps the view instantly.
- On a computer you can watch a screen instead of a camera.
- Auto screen capture adapts to the display, with a freeze-and-teach HUD.
- On macOS, Cmd+Shift+B grabs a region into the open project.
- Front and back switching works on phones.
- The camera keeps working while Belo retrains.
Hold to teach
Tap to teach one picture. Hold and they keep coming, up to 60. Try it.
2 to 12 pictures a second, your pick. If the disk falls behind, pictures get dropped.
Find your own objects with real boxes
Choose Detect objects when you start a project. Draw boxes around the things you care about, press Train, then use the camera. Belo can find your custom objects and draw a box around each one. It is not limited to a list of common objects someone else chose.
About 50 to 200 boxes for each label usually helps. Training on desktop is recommended.
After a while, Belo draws the boxes for you
Box a few pictures yourself. Belo trains on them in the background, and from then on it opens each new picture with its guesses already drawn on it, dashed, with a name and how sure it is. Click a guess to keep it, press its x to throw it away, or just draw your own box straight over it.
Keep all of them with one press of enter. Press space instead and Belo keeps them and opens the next picture, so a folder of photographs becomes a row of yes, yes, yes rather than an afternoon of dragging rectangles.
Made up numbers. A guess is never saved into your project until you keep it, and leaving the picture throws away the ones you did not.
Borrowed eyes, a small brain of your own
Eyes
MobileNet v1, 224 pixels
A ready made picture reader, built in. It turns pictures into clues, and never phones home.
Brain
100 units, seconds per run
The only part that learns your labels. That is why training takes seconds, not hours.
Engine
TensorFlow.js, graphics card first
Graphics card if possible, processor if not. No server, no API key, no sign up.
Undo
One spare copy, one button
Belo copies the old model before retraining itself. One button brings it back.
A project is a folder you can open
A plain folder named Project.belo. Same on Mac, Windows and Linux. Open it, copy it, back it up.
Belo writes a temp file and renames it, so a crash cannot leave half a file. When a trained model stops matching your labels, its files are deleted rather than left lying around.
- One small file holds labels, pictures and settings.
- Deleting asks first, and only ever touches real projects.
- Export model saves one zip that runs outside Belo: the trained part, the extractor it reads, a description of the format and a page that runs it. It is a developer bundle, not a phone app.
- On iPhone, moving projects means a device backup for now.
Shapes.belo/
project.json your labels, pictures and settings
images/
img_a1f.jpg one picture, trimmed to 224x224
img_b73.jpg
img_c02.jpg
model/
head.json what classification learned
detect.json what object detection learned
head.prev.json the older copy, for undoGet it
Download Belo.
Free and open.
Pick your machine. Every installer is built by GitHub Actions from the tagged source, and each new version keeps the same links.
macOS
Download the .dmgThe build is not signed by Apple yet. The first time, right click Belo in Applications and choose Open, then confirm. A double click alone will be blocked.
Windows
Download the installerPrefer an msi? Belo-windows-x86_64.msi
The build is not signed yet, so SmartScreen may warn you. Choose More info, then Run anyway.
Linux
Download the AppImageOn Debian or Ubuntu: Belo-linux-x86_64.deb
Mark the AppImage executable before the first run. The camera depends on your system, and Linux is the least tested.
See every file in the latest release
Or build it from source
The installers come from this same source. If you would rather build it yourself, you need Node 20, Rust 1.77 and the Tauri 2 system packages.
git clone https://github.com/isaaclins/belo.git
cd belo
npm install
npm run tauri build- macOS
- Windows
- Linux
A source build is unsigned too, and it is the only way to get an Intel Mac build today.
Open the repo on GitHubiPhone and iPad
App Store
Coming soon, pending Apple review
The iPhone build works today. Apple is still reviewing the developer license, so there is nothing to link yet.
Needs iOS 15. The simulator has no camera, so you add photos instead. Android is written but not built.