Autodistill lets you label images automatically using foundation vision models [1], in one command. You can use this data to train smaller, fine-tuned models which require significantly less compute to run.

Distilled models are both interpretable (you know what data went into training), and require less compute resources to run.

Autodistill makes it easier to get to an MVP that uses your data for use in testing vision. Then, you can refine your annotations and model to the extent required for production.

For example, I trained a milk bottle detection model in a few commands (one for trying prompts, another for labeling data and training a model) [2]

    [1] https://blog.roboflow.com/ai-vs-human-labeled-data/
    [2] https://blog.roboflow.com/label-train-deploy-autodistill/