autodistill

Why we included this project

Getting a custom detector off the ground usually means hand-labeling thousands of images, which is where most vision projects slow to a crawl. Autodistill skips that step by running a large foundation model as a teacher over your unlabeled images and using the labels it produces to train a smaller supervised model, one light enough to run at the edge. The plugin API keeps the pieces swappable, so Grounding DINO or Segment Anything can act as the teacher while YOLOv8 or DETR serve as the target detector, and switching models does not require reworking your pipeline. That tradeoff is the point: you pay for one expensive labeling pass up front and get cheap, low-latency inference afterward. If you know the classes you want to detect but do not have a curated dataset, Autodistill takes you from raw images to a deployable model without a labeling tool in between.