It uses a lightweight ML model to calculate image similarity (cosine similarity of the image embeddings), so it will work on reencoded jpegs and such. The similarity threshold for detecting duplicates is configurable, so you can play around with it if the default doesn't work for you.
It also displays file size & dimensions when reviewing duplicates, but doesn't use that to determine similarity.
It also displays file size & dimensions when reviewing duplicates, but doesn't use that to determine similarity.