Install
One line.
pip install fotonetEverything Studio trains with is in the library.
fotonet Studio
Label. Train. Ship.
All on your desktop.
Coming soon to Windows, macOS and Linux.
fotonet Studio
Six tabs.
One whole workflow.
fotonet Studio is our desktop app for object detection. Label your images, train a fotonet model on them, test it and export it, in one window. It runs offline, and your data stays on your machine.
Boxes, fast. Image after image.
The Label tab is where a dataset is made. The image list on the left shows what is done, what is still to do and which images are backgrounds. Draw a box, pick its class, and press Save and next.
- Draw, move and resize boxes on the photo itself, with a shortcut for everything.
- The magic wand finds an object from one click.
- Auto-label lets a trained model draft the boxes for a whole folder. You review them.
- Deleted images and labels can be restored.
See it before you train on it.
Augmentation makes new training images from the ones you have: flipped, shifted, distorted, recoloured. The Augment tab shows the original and the result side by side, boxes included, so you know exactly what the model will see.
- Pick filters from a searchable list and stack them in order.
- Each filter has its own strength and chance of being used.
- The preview follows the boxes, so a bad transform shows at once.
- Generate extra copies of every image when you want them on disk.
Watch it learn.
Pick a size, start from scratch or from released weights, and press one button. Every setting of the trainer sits on the left, and a dot marks what you changed. Loss, accuracy, speed and the machine itself draw live while it runs.
- Nano, Small, Medium and Large, with the same settings for each.
- Start from scratch, from pretrained weights, or resume a stopped run.
- Live charts for every loss and for mAP50-95 and mAP50.
- Each run writes its full recipe, so the same run is one command line away.
Every run, side by side.
Results keeps every run of the project with its curves and its best scores. Pick a checkpoint and compare what it predicts with your own labels on any validation image, at any confidence.
- Best mAP50-95, mAP50, precision and recall per run.
- Ground truth and prediction next to each other, image by image.
- Use a checkpoint to auto-label, or fine-tune it further.
- Export to ONNX, TorchScript, TensorRT or CoreML.
Try it on anything.
Point any checkpoint at an image, a folder, a video file or your camera and see what it finds, with a count for every class and the time it took.
- Image, folder, video and camera sources.
- One confidence slider. No NMS threshold, because there is no NMS.
- A count of everything found, class by class.
- Save the annotated result.
Know your dataset.
The Data tab measures the dataset before you train on it: how many images and boxes, how balanced the classes are, how big the objects are and where they sit in the frame.
- Boxes and images per class, with a warning when classes are unbalanced.
- Box sizes in the COCO small, medium and large bands.
- A heat map of where the boxes are.
- Health checks, video frame import and backups in their own views.
Label
Boxes, fast.
Image after image.
Draw, move and resize boxes on the photo itself. Zoom to the pixel, pan anywhere, and let the magic wand or a trained model draft the boxes for you to check.
-
Draw and edit. Every box moves and resizes in place, with a shortcut for everything.
-
Magic wand. Click an object and get a box around it.
-
Auto-label. A trained model drafts the boxes for a whole folder. You review them.
-
Nothing is lost. Deleted images and labels can be restored.
Train
Watch it learn.
Know where it's going.
Start a run with one button. Every loss and every metric draws live, with a forecast of where the run will land and guard rails that flag a run going wrong.
Augment, then train.
See exactly what every augmentation does to your images before a run uses it.
Stop. Resume. Compare.
Stop a run whenever you need the machine, resume it later from where it stopped, and compare every run you have made.
Test and export
Try it on anything.
Ship it anywhere.
Run any of your checkpoints on an image, a folder, a video or your camera. Compare its boxes with your labels side by side, and export it when it's ready.
Test on
- An image
- A folder
- A video
- Your camera
Export to
- ONNX
- TensorRT
- TorchScript
- CoreML
Agents
Ask your agent.
Watch it label.
Tell your AI agent what you need in plain words. Agents label your images, fix boxes and tidy your classes in fotonet Studio, and you watch every box land.
Label every car, person and bike in my parking lot photos.
-
A whole folder at once. Agents split the work and label side by side. Intelligently.
-
Nothing without your OK. You decide what an agent may do on its own. For the rest, Studio asks first, and you can say yes, no, or what to do instead.
-
See every step. Studio opens each image as the agent works on it and takes on its model's colour, so you see the boxes as they land.
-
Undo anything. Every change an agent makes is logged and can be undone. Removed images go to the trash.
fotonet Studio's MCP server is tuned for Claude models and works best with them. Claude is a trademark of Anthropic, PBC; fotonet is not affiliated with Anthropic.
Coming soon.
Start with the library today.
Train
From the command line.
fotonet train \
model=fotonetn \
data=data.yamlFollow along
Watch the repository.
Studio's release will be announced on GitHub.