aiiu-lab/DFD-FCG ? reverse-engineered prompt

Build me a deepfake video detection project like this paper, where I can train a model, evaluate it on different datasets, and run inference on a single video or a demo clip.

I want the app to take preprocessed face aligned videos, detect whether a video is real or fake, and work with the provided dataset structure and scripts for training, testing, and robustness checks. Please include the preprocessing flow for extracting face landmarks, cropping the main face, and handling the robustness version too. Also make sure there is a simple inference path from the command line so I can point it at one video and get a prediction.

If there are any missing details from the paper or current setup, look up the latest docs or whatever is needed to make the project actually run. Keep it practical, with clear commands for setup, training, evaluation, and demo usage.