This repository implements a deep learning model for face swapping in images and videos, based on the deepfakes autoencoder architecture extended with adversarial losses and attention mechanisms. It uses a U-Net style denoising autoencoder with self-attention (SAGAN) and perceptual loss computed from a VGGface network. The system supports multiple output resolutions (64x64 to 256x256) and includes tools for face detection and alignment via MTCNN. Training is done in Jupyter Notebooks, with Google Colab support provided.