Apps
Published bundles, Hugging Face, the one-line CLI, 3D Slicer.
What an app is
A KonfAI app packages a trained model, its weights and its configuration: a folder with app.json, the Prediction.yml (and the train and evaluation configs when the app ships them), the checkpoints, and the Python code the model needs. It is the same contract everywhere: a local folder, a Hugging Face repository (repo_id:app_name) or an HTTP server.
Run one
python -m pip install "konfai[imaging]" konfai-apps
konfai-apps infer VBoussot/ImpactSeg:body -i ./scan.mha -o ./Output --gpu 0Replace ./scan.mha with your image; --cpu 1 runs on the CPU. The first run downloads the app's weights and installs any missing app requirements. An app imports Python code and installs its requirements by default: use a source you trust (KONFAI_APPS_INSTALL_REQUIREMENTS=0 to manage requirements yourself). The output lands under Output/<app>/Output/P000/; open the label image beside your input in a viewer such as 3D Slicer and check that they align.
The konfai-apps command also evaluates an app against a ground truth, stacks test-time copies into uncertainty maps, fine-tunes an app that ships a train config, and bundles a trained session into a new app. From Slicer, the SlicerKonfAI and SlicerImpactReg extensions run the same apps.
The published apps
The catalogue on this site: Apps (segmentation, synthesis and registration bundles with their measured figures). The task guide: Run a published app; packaging your own: Package an app.