Open-vocabulary object detection

You describe what to look for in words and it is found, with no retraining and no labelled dataset to assemble first. That makes it the fastest way to test a new question against existing imagery before committing to a trained model. Where a description proves useful, it can be promoted into a trained model later.

You can use the models we operate or bring your own model and we run it on your cluster. Work arrives as an API call, a batch over a folder or bucket, or a resident on-premise service, and results return as structured output with a per-job audit record.