Comment by nicklewers

Thanks! So yeah its all on-device using Apples Vision framework. I use feature prints and embeddings + face counts + person counts + a few other properties to cluster similar photos, then score each one based on sharpness, exposure, face clarity, composition, emotional score, and Apple’s aesthetic score. Each of those is weighted a bit differently.

You don’t need to define “similar” yourself, you select i.e. 100 photos from your library, then it figures that out the groups by itself, and there’s the review screen so you can confirm the decisions made before deleting anything.