Key Features

Face Swap Components and Architecture Overview

AHFM: Advanced Hyper Face Model

A deep learning architecture designed for high quality face swapping in images/videos. It features an encoder/decoder structure, incorporating Generative Adversarial Networks (GANs) and temporal consistency mechanisms.

Features:

• Face Swapping: Combines identity preservation, style modulation, and blending for realistic face swaps.
• Video Support: temporal consistency through the Temporal Consistency Module and Temporal Discriminator.
• Progressive Growing: Training begins at a lower resolution (256) and progressively increases to the target resolution (1024) for improved stability.
• Style Modulation: Utilizes ModulatedConv2d for adaptive style injection.
• Data Augmentation: Includes random flips, warps, color adjustments, and mask filters.
• Adversarial Training: Integrates spatial and temporal discriminators with configurable GAN loss weights.
• Pretraining and Optimization: Supports the use of pretrained weights.
• Model Optimization: Supports INT8 quantization for faster inference and reduced model size.
• ONNX and PTH Export
• Flexible Configuration: Extensive hyperparameter options for resolution, filter sizes, augmentation, and loss weights.
• Distributed Training: Supports Distributed Data Parallel (DDP) for multi-GPU training.
• Segmentation Integration: integrates GSM for face mask generation.

The AHFM model architecture supports three resolution types:

• 256, 512 and 1024.

GSM: Genome Smart Mask

• A smart masking system designed to determine and isolate only the region that needs to be swapped.

GFD: Genome Face Detector

• An advanced face detection and landmark system optimized for complex scenes, small faces, and multiple faces. It maintains high accuracy under varying lighting conditions and image quality.