Model training for edge video analytics has to balance cloud-scale representational power with strict edge-side constraints on bandwidth, energy, and privacy. The problem can be tackled through three architectural paradigms: cloud-centric, edge-native, and hybrid edge-cloud. Cloud-centric training leverages elastic GPU pools to optimize large CNNs, Transformers, and generative models on centrally collected video, but incurs heavy uplink traffic and privacy risks. Edge-native training avoids raw-data uploads through decentralized federated learning. Hybrid edge-cloud computing paradigm exploits both edge and cloud computing advantages, striking the balance among the performance of accuracy and of latency, etc.