I am using BlazeFace for face detection and MobileFaceNet for face recognition. Based on my experiments, accuracy is lower when comparing embeddings across different devices.

I just want to know theoretically whether cross-device face embedding comparison is possible, assuming the same model and pipeline are used.

For example, Device A generates and stores a face embedding. Later, Device B uses the same model and pipeline to generate another embedding and compare it using cosine distance.

Assuming both devices use the same detection model, recognition model, image preprocessing, input size, normalization, and embedding comparison logic, should the embeddings be compatible across devices?

In other words, can an embedding generated on one device be reused for nearest-neighbor or cosine-distance matching on another device?