The vision model
studio.
LAI is the self-hosted way to annotate datasets and train computer vision models — built for engineers, MIT licensed.
// platform
Everything between data and a trained model.
Projects group datasets, models, evaluations and exports. One self-hosted stack, no glue code.
Datasets
Create, merge and augment datasets. Chunked uploads for images, videos and collections with tags, classes and class colors.
Assisted Annotation
SAM-powered segmentation in an image viewport with toolbar, minimap, zoom and COCO import.
Auto-Annotate
Run any trained model over a dataset to bootstrap labels, then refine with humans in the loop.
Training
Train popular vision models (e.g. YOLO, RT-DETR, RF-DETR) on your own GPUs — or bring your own architecture.
Evaluation
Confusion matrices, threshold explorer and side-by-side evaluation comparison to find failure modes.
Export & Inference
Export trained models and test inference directly in the studio before shipping.
// workflow
From install to edge deployment in 6 steps.
Install
Python CLI wrapping Docker Compose. One command and the studio is up.
$ pip install -e . && lai install-guiRun
Bring the full stack up locally on port 8089.
$ lai upAnnotate
Upload images, videos or COCO and label with SAM-assisted tools.
$ open http://localhost:8089Train
Pick a popular vision model (e.g. YOLO, RT-DETR, RF-DETR) or add your own, then launch a run on your GPU.
$ # from the studio: Train modelEvaluate
Inspect confusion matrices, sweep thresholds and compare runs.
$ # from the studio: EvaluateExport
Convert checkpoints to ONNX, TensorRT, CoreML or TFLite and ship to your edge device.
$ # from the studio: Export → ONNX / TRT// the studio
A look inside.
Organize every project
Keep datasets, models and experiments grouped per project so teams can move between problems without losing context.
Curate your datasets
Import, version and inspect datasets in one place — splits, class distributions and sample previews always one click away.
Train on your own GPUs
Configure popular vision models (e.g. YOLO, RT-DETR, RF-DETR) — or plug in your own — and watch live loss, mAP and GPU utilization stream into the studio.
Evaluate and compare
Confusion matrices, threshold sweeps and side-by-side prediction vs ground-truth views make failure modes obvious.
Export anywhere
Convert trained checkpoints to ONNX, TensorRT, CoreML or TFLite and ship them straight to your edge or cloud runtime.
Train your next vision model.
Open source, MIT licensed, and runs entirely on your own hardware.