Autodistill

Autodistill is an open source Python library from Roboflow that uses large, slow foundation models to train small, fast supervised models for computer vision. You feed it unlabeled images, a large base model labels them automatically according to an ontology you define, and the resulting dataset trains a compact target model that can run at the edge. The aim is to go from raw images to inference on a custom model with no human labeling in between. It is built for machine learning engineers and developers who need a domain specific detector or classifier without paying for manual annotation. The library defines clear primitives: a task, a base model such as GroundedSAM, an ontology that maps prompts like text captions or CLIP vectors to class names, the auto labeled dataset, a target model such as YOLOv8 or DETR, and the distilled model that comes out. The interface is pluggable, so base and target models can be swapped, and the documentation covers object detection, instance segmentation, image classification and text classification along with community contributed models. You keep ownership of the trained model and can deploy it to the cloud or to edge hardware. Autodistill is free under the Apache 2.0 license and runs on your own hardware. Roboflow also offers a hosted version that labels images in the cloud; its pricing sits on Roboflow's platform and is not published in the Autodistill documentation. The usual trade off applies: distilled target models are fast and accurate on the classes they were trained for, but the documentation notes they do not generalize well beyond their dataset.