Deep-Live-Cam: Real-Time Face Swapping With Just One Photo

By Prahlad Menon Published 2026-04-21 4 min read

What if you could swap your face in a live video call — in real time — using nothing but a single photograph?

That’s exactly what Deep-Live-Cam does. And with over 91,000 GitHub stars and 13,000+ forks, it’s become one of the most popular open-source AI projects in the world.

What Deep-Live-Cam Actually Does

Deep-Live-Cam performs real-time face swapping in video streams. You provide one source image (the face you want to wear), point it at a webcam or video feed, and it swaps your face — live, in real time. No model training. No dataset collection. One image, three clicks.

The core pipeline:

Face detection — Identifies faces in each video frame
Face swap — Uses the InsightFace/inswapper model to replace detected faces with your source image
Face enhancement — GFPGAN restores quality and detail post-swap
Frame output — Renders the result to a virtual camera or output stream

The entire process runs at interactive framerates on consumer hardware — NVIDIA GPUs via CUDA, Apple Silicon via CoreML, or even CPU-only (slower but functional).

Why This Project Matters

Deep-Live-Cam sits at a fascinating intersection of capability and accessibility. A few years ago, creating a convincing face swap required:

Hundreds or thousands of training images
Hours of model training on expensive GPUs
Significant technical expertise
Post-processing and manual cleanup

Now it takes one photo and a few seconds of setup. The democratization curve here is steep and worth understanding.

The Technical Stack

Under the hood, Deep-Live-Cam combines several well-established models:

Component	Model	Purpose
Face Swap	inswapper_128 (InsightFace)	Core face replacement
Face Enhance	GFPGANv1.4	Post-swap quality restoration
Face Detection	InsightFace buffalo_l	Real-time face detection
Execution	ONNX Runtime	Cross-platform inference

The ONNX Runtime backend means it runs across CUDA (NVIDIA), CoreML (Apple Silicon), DirectML (AMD/Intel on Windows), and plain CPU — making it genuinely cross-platform.

Key Features

Mouth Mask — Retains your original mouth movements for natural-looking speech
Face Mapping — Apply different faces to multiple people in the same frame simultaneously
Live Streaming — Output to virtual cameras for use in Zoom, Teams, OBS, or any video app
Video Processing — Process pre-recorded videos with face swaps
Movie Mode — Watch movies with any face swapped in, in real time

The Ethics Question

The project includes built-in safeguards — NSFW content detection that blocks processing of inappropriate material. But the ethical dimensions of real-time deepfakes are significant:

Legitimate uses:

Content creators animating custom characters
Virtual try-on for fashion and entertainment
Privacy protection in video calls
Film and media production on a budget
Accessibility — letting people with facial differences participate in video without anxiety

Risks:

Identity fraud and impersonation
Non-consensual deepfake creation
Erosion of trust in video as evidence
Potential for harassment and scams

The project maintainers are explicit: if you use a real person’s face, get consent and label outputs as deepfakes. The AGPL-3.0 license also ensures all derivative works must remain open source.

How It Compares

Deep-Live-Cam isn’t the only face-swapping tool, but it’s arguably the most accessible real-time solution:

Tool	Real-Time	Training Required	Open Source	Stars
Deep-Live-Cam	✅	None (1 image)	✅ AGPL-3.0	91K+
DeepFaceLab	❌	Hours of training	✅	48K+
FaceFusion	✅	None	✅	25K+
Roop	❌	None	✅ (archived)	26K+

Deep-Live-Cam essentially evolved from the Roop project (now archived) and added the critical real-time capability that makes it practical for live use cases.

Getting Started

The simplest path:

git clone https://github.com/hacksider/Deep-Live-Cam.git
cd Deep-Live-Cam
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Download the two required models (GFPGANv1.4 and inswapper_128_fp16.onnx) from HuggingFace and drop them in the models/ folder. Then:

python run.py                          # CPU mode
python run.py --execution-provider cuda  # NVIDIA GPU

For Apple Silicon users, install onnxruntime-silicon for CoreML acceleration.

The project also offers pre-built binaries for Windows and Mac if you’d rather skip the manual setup.

The Bigger Picture

Deep-Live-Cam is a signal of where real-time AI is heading. The models are getting smaller, the inference is getting faster, and the barrier to entry is approaching zero.

What took a research lab in 2019 now runs on a laptop in 2026. The face-swapping capability itself is table stakes — the real question is what the next generation of real-time video manipulation looks like, and whether our detection and verification systems can keep pace.

For now, Deep-Live-Cam remains a remarkable piece of engineering: a single-image, real-time, cross-platform face swapper that anyone can run. Use it responsibly.

📍 Source: github.com/hacksider/Deep-Live-Cam