The performance is focused on correctness and the APIs work with individual images. The underlying models can be run with batches, but you need to extract the architecture code to run directly. As a result, while I started with DeepFace, I mostly just used the models.