Can I ingest embeddings instead of frames for training, labelling and predictions for the AWS Rekognition custom label?

I am thinking of storing image embedding in a feature store and use the AWS Rekognition custom label and I am weighing whether it is better to store frames/embeddings/image hashes. But before that I would like to know if it is possible to use embeddings/image hashes for training/prediction/labelling rather than images. Thank you!