davidsandberg/facenet ? reverse-engineered prompt

Build me a face recognition project in Python using TensorFlow that can detect, align, and compare faces from images.

I want it to be able to train a model on face datasets, create face embeddings, and use those embeddings to recognize whether two photos are the same person. It should also include face alignment with MTCNN, a way to run validation on a benchmark like LFW, and support using pretrained models for inference right away. If training is included, make it easy to train either a classifier or an embedding model, and keep the scripts simple enough that I can point it at my own image folders and get results.

Please make the project work cleanly end to end, with clear setup instructions, working tests if possible, and sensible defaults for preprocessing like image standardization. If you need to check current TensorFlow or model usage details, look them up online first.