video-face-swap

Processes video frames sequentially to detect and replace faces while maintaining temporal coherence across frames. Uses deep learning models (likely DeepFaceLab or similar face-swap architecture) to extract facial embeddings from a source face, then applies morphing and blending operations to target video frames. The Gradio interface handles video upload, frame extraction, model inference batching, and video reconstruction with audio preservation.

Detects facial landmarks in both source and target video frames using a face detection model (likely MTCNN, RetinaFace, or similar), extracts facial embeddings via a pre-trained encoder (e.g., FaceNet, ArcFace), and computes geometric alignment matrices to warp the source face to match target head pose and scale. This alignment step ensures the swapped face fits naturally into the target frame's spatial context.

After face alignment, applies pixel-level blending operations (e.g., Poisson blending, alpha blending with feathered masks) to seamlessly merge the warped source face into the target frame. Includes color histogram matching or adaptive color correction to reduce visible seams and ensure the swapped face matches the target frame's lighting, skin tone, and color temperature. Operates on each frame independently to avoid temporal flickering.

Automatically extracts all frames from input video at the original frame rate using FFmpeg, processes them through the face-swap pipeline in batches (leveraging GPU parallelism), and reconstructs the output video by encoding processed frames back to MP4 with H.264 codec while preserving the original audio track. Handles variable frame rates and resolutions transparently.

Provides a Gradio interface that handles file uploads, manages inference queue, displays progress, and serves downloadable results. Gradio abstracts away model loading, GPU memory management, and HTTP request handling, allowing the face-swap pipeline to be exposed as a simple web form with file inputs and a download button. Runs on HuggingFace Spaces infrastructure with ephemeral GPU allocation.