Deep-Live-Cam Hits 83K Stars Swapping Faces in Real Time With One Photo
Deep-Live-Cam hits 83K GitHub stars: one photo is now all it takes to swap your face in real time on any webcam or video stream
Takeaways
- Deep-Live-Cam is a free, open-source Python tool that swaps faces in real-time webcam or video streams using just one reference photo and no model training.
- The pipeline chains InsightFace (detection + alignment),
inswapper_128(GAN-based swap), and GFPGAN (face restoration) entirely on-device with no cloud uploads. - Supports NVIDIA CUDA, Apple Silicon CoreML, AMD/Intel DirectML, and OpenVINO backends; hits 60+ FPS on an RTX 4090, 20-30 FPS on mid-range consumer GPUs.
- Key features include Mouth Mask (preserves real lip movement), multi-face mapping, Many Faces mode, and a built-in NSFW content filter.
- The tool has crossed 83K GitHub stars and is directly linked to a surge in deepfake-based identity fraud, with deepfaked selfie attempts rising 58% in 2025 per Entrust's 2026 report.
- Available free via GitHub; a pre-built quickstart with 30+ extra features is sold at deeplivecam.net.
Deep-Live-Cam has quietly become one of the most-starred AI repositories on GitHub, crossing 83,000 stars with over 12,000 forks. The premise is almost absurdly simple: give it one photograph of a face, point it at your webcam, and your face is replaced in real time. No training run, no dataset, no cloud API. Just a Python script and a consumer GPU.
One photo, zero training
Deep-Live-Cam is a real-time face-swapping system that lets you replace faces in live video with a single photograph, instantly, without training datasets or waiting hours for model convergence. That last part is the real unlock. Previously, face-swap technology required GPU clusters, hours of training data, and specialist knowledge. Deep-Live-Cam collapses that barrier entirely: you need one photo and about five minutes to get started.
Deep-Live-Cam uses a fundamentally different approach: single-shot inference with pre-trained models. Instead of training a custom model for each face, it leverages the inswapper_128 model, a pre-trained face-swapping neural network that can generalize to any face from just one image. Think of it as a universal face translator that already speaks every face's language.
How the pipeline actually works
The technical stack combines three well-established components: InsightFace (a face detection and analysis model that locates and aligns faces in each video frame), GFPGAN (a face restoration neural network that sharpens and corrects the swapped face so it looks natural), and ONNX Runtime (an inference engine that runs neural network models efficiently without requiring the original training framework).
Here is what happens on every single frame:
- InsightFace's detection models (RetinaFace or YOLO-based detectors) locate all faces in each frame and extract 106 facial landmarks for alignment.
- The system generates 512-dimensional face embeddings for matching, then applies the source embedding to the target face via the
inswapper_128.onnxgenerator model. - The model infers 3D facial structures from 2D images and separates identity features from pose features, which allows for smooth face replacements.