Clustering face embeddings (512d) using GCN's (not knowing the amount of needed clusters)

Follow the full discussion on Reddit.
Hi, I have used the InsightFace model to detect a bunch of faces in images. It returns face embeddings of 512d. I don't have reference data for the persons on these images, neither do I know how many different identities appear on them. I would love to cluster the face embeddings as good as possible. So far I have tried dbscan and hierarchical clustering, both showing decent results when I manually evaluate the clusters (after playing around with the hyperparams).