Thanks for your response. I should clarify that I’m still at the very beginning of this project, so I don’t yet have a full codebase or dataset to share. My current experiments are with ArcFace/MobileFaceNet in PyTorch, mainly to understand the pipeline end‑to‑end (detection → alignment → embedding → matching). My main goal is to support incremental enrollment of new identities without retraining the whole model from scratch. In other words, I’d like the system to adapt over time while minimizing catastrophic forgetting and avoiding heavy retraining cycles.

I understand fine‑tuning is a common approach, but I’m trying to figure out what architecture choices make this easier. For example:

• Should I stick with a frozen embedding backbone (ArcFace/MobileFaceNet) and just store new embeddings in a gallery (prototype‑based)?

• Or should I look into parameter‑efficient methods (adapters/LoRA) or knowledge‑distillation strategies to update the model incrementally?

• Are there recommended backbones (ResNet, Vision Transformer, etc.) that are better suited for continual learning in face recognition?

Since I’m still learning, I’d appreciate guidance on which model architecture is most practical for a project like this, where retraining from scratch isn’t feasible. My aim is to build something simple first (incremental enrollment) and then explore more advanced continual learning methods if time allows.