Autodistill

Autodistill automates the labeling step of computer vision training: a large foundation model such as Grounding DINO, Grounded SAM 2, OWLv2, or DETIC labels raw images according to a user-defined ontology of text prompts, and the resulting dataset trains a small target model like YOLOv8, YOLO-NAS, or DETR that is cheap enough for edge deployment. The outcome is a supervised detector or segmenter produced with no manual annotation, following a distillation pattern where slow general models teach fast specialized ones. The framework is modular, with each base and target model in its own pip installable plugin package, plus connectors for cloud vision APIs from Azure, AWS, and Google. Everything is Python, installed with pip install autodistill plus the plugins a project needs, and the heavy base models run best with a CUDA GPU although CPU inference works. Roboflow released it under the Apache 2.0 license with documentation at docs.autodistill.com; the core package has been stable since 2025 while newer plugins such as SAM 3 and RF-DETR ship separately. It holds around 2,800 GitHub stars.