Deep-Live-Cam — Run Real-Time Face Swap on Your Own PC

Full local setup, plus a prompt that makes an AI agent install it for you

What is Deep-Live-Cam?

Real-time face swap on your webcam, running on your own machine

From nothing to a live face swap

1. Windows or macOS + webcam — GPU optional, ~10x faster
2. How do you want to install it?
   • Let an AI agent do it — paste the prompt above
   • Do it by hand — steps 1–6, ~20 min
3. Install the prerequisites — Python 3.14, git, ffmpeg
4. Clone the repository
5. Create a virtual environment and install — the slow part
6. Download the two models — check the file sizes
7. Turn on GPU acceleration — skip on CPU
8. Run it
9. Which command for your hardware?
   • NVIDIA GPU — cuda
   • Apple Silicon — coreml
   • CPU only — slow but works
10. Pick a face, go live — first launch pulls ~300MB

Deepfake software — only swap faces you have permission to use.

[Deep-Live-Cam](https://github.com/hacksider/Deep-Live-Cam) swaps a face onto your live webcam feed from **a single photo** — no training, no dataset. It also does image and video face swaps offline.

• **Runs entirely on your computer** — nothing is uploaded anywhere
• **One source photo** is all it needs
• **Live webcam mode** — pair it with OBS to use in calls or streams
• **Free and open source**, and it works on CPU (just slower)

Read this before you install This is deepfake software. The project ships a built-in filter that blocks nudity and graphic content, and its licence asks you to **get consent** before using a real person's face and to **label output as a deepfake** when you post it. Using someone's face to impersonate them — on a video call, to open an account, or to make it look like they said something — is fraud in most countries. Use your own face, or a face whose owner said yes.

Before you start

• Python 3.14 recommended (3.11–3.14 all work — NOT 3.10, onnxruntime dropped it)
• git and pip
• ffmpeg
• Windows: Visual Studio 2022 C++ runtimes
• macOS Apple Silicon: python-tk (the GUI needs tkinter)
• NVIDIA GPU optional — CPU works, just slower
• ~2GB of disk for the repo, dependencies and models

The lazy way: let an AI agent install it

Paste this into Claude Code, Codex, Cursor or Gemini CLI

If you have a coding agent installed, you can skip the manual steps. Open it **in a terminal**, paste the prompt below, and let it do the install. It figures out your OS and GPU and picks the right commands.

Install Deep-Live-Cam on this computer and get it running.

Repo: https://github.com/hacksider/Deep-Live-Cam

Do this:
1. Detect my OS (Windows / macOS Apple Silicon / Linux) and whether I have an
   NVIDIA GPU. Tell me what you found before you install anything.
2. Check for Python 3.11-3.14 (3.14 preferred), pip, git and ffmpeg. Install
   whatever is missing using the right package manager for my OS. On Windows
   also check for the Visual Studio 2022 C++ runtimes. On macOS also install
   python-tk (the GUI needs tkinter).
3. Clone the repo into a sensible folder and cd into it:
   git clone --depth 1 https://github.com/hacksider/Deep-Live-Cam.git
4. Create a venv with that Python version, activate it, and
   pip install -r requirements.txt
5. Create a `models` folder if it does not exist and download both model files
   into it:
   - https://huggingface.co/hacksider/deep-live-cam/resolve/main/GFPGANv1.4.onnx
   - https://huggingface.co/hacksider/deep-live-cam/resolve/main/inswapper_128_fp16.onnx
   Verify each file downloaded fully (they should be tens/hundreds of MB, not a
   few KB of HTML error page).
6. If I have an NVIDIA GPU, set up CUDA acceleration:
   pip install -U torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
   pip uninstall -y onnxruntime onnxruntime-gpu
   pip install onnxruntime-gpu==1.21.0
   Then the run command is: python run.py --execution-provider cuda
   On Apple Silicon use: python run.py --execution-provider coreml
   Otherwise plain CPU: python run.py
7. Launch it once to confirm the GUI opens, then stop and tell me the exact
   command to start it again next time.

Rules:
- Show me each command before you run it.
- Do not modify anything outside the Deep-Live-Cam folder except installing the
  listed system packages.
- If a step fails, show me the actual error and your fix before continuing.
- At the end, print a short summary: install path, Python version, GPU or CPU
  mode, and the start command.

Which agents does this work in? Anything that can run terminal commands: **Claude Code**, **OpenAI Codex CLI**, **Cursor** (agent mode), **Gemini CLI**, or **Windsurf**. A plain chatbot without terminal access can't install anything — it will only read the commands back to you.

Watch what it runs The prompt tells the agent to show each command before running it. Actually read them. An agent installing system packages has real reach over your machine, and you are the one approving it.

1. Install the prerequisites

Python, git, ffmpeg — and the C++ runtimes on Windows

Get **Python 3.14** from [python.org](https://www.python.org/downloads/) (tick *Add Python to PATH* on Windows), plus [git](https://git-scm.com/downloads) and [ffmpeg](https://ffmpeg.org/download.html).

iex (irm ffmpeg.tc.ht)

Windows also needs the C++ runtimes Install the [Visual Studio 2022 build tools](https://visualstudio.microsoft.com/visual-cpp-build-tools/). Without them the dependency install fails with compiler errors.

brew install [email protected] [email protected] ffmpeg git

2. Clone the repository

A shallow clone is enough

git clone --depth 1 https://github.com/hacksider/Deep-Live-Cam.git
cd Deep-Live-Cam

3. Create a virtual environment and install

Keeps these packages away from your system Python

python -m venv venv
venv\Scripts\activate
pip install -r requirements.txt

python3.14 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Takes a few minutes It pulls torch and onnxruntime, which are large. A slow first run is normal.

4. Download the two models

They go in the models folder

Create a `models` folder inside `Deep-Live-Cam` and put both files in it: 1. [GFPGANv1.4.onnx](https://huggingface.co/hacksider/deep-live-cam/resolve/main/GFPGANv1.4.onnx) — face restoration 2. [inswapper_128_fp16.onnx](https://huggingface.co/hacksider/deep-live-cam/resolve/main/inswapper_128_fp16.onnx) — the face swapper

Check the file sizes If a download finishes instantly and the file is a few KB, you saved an error page instead of the model. Both files should be tens to hundreds of MB. This is the single most common reason it fails to start.

5. Turn on GPU acceleration

Optional — skip if you don't have an NVIDIA card

CPU works but is slow for live mode. With an NVIDIA GPU, install [CUDA Toolkit 12.8](https://developer.nvidia.com/cuda-12-8-0-download-archive) and cuDNN v8.9.7 for CUDA 12.x, then:

pip install -U torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
pip uninstall onnxruntime onnxruntime-gpu
pip install onnxruntime-gpu==1.21.0

6. Run it

Pick the command that matches your hardware

python run.py

python run.py --execution-provider cuda

python3.14 run.py --execution-provider coreml

First run downloads more models Roughly 300MB extra on the first launch. Let it finish.

Things other tutorials get wrong

Checked against the live repo on 26 August 2026

Python 3.10 does not work any more Most guides still tell you to install Python 3.10. That was true once — it isn't now. **onnxruntime dropped 3.10**, so the current project needs **3.11–3.14, with 3.14 recommended**. If you followed an older tutorial and the dependency install died, this is why.

It's AGPL-3.0, not MIT Deep-Live-Cam is licensed **AGPL-3.0**. Almost nobody mentions this. It matters if you plan anything commercial: AGPL means that if you run modified code as a network service, you have to publish your source. Personal and content use is fine. Building a paid product on it is not something you can do quietly.

Deep-Live-Cam vs the browser face swappers

Pick based on what you're actually doing

Using it

Two modes

1. **Image / video** — pick a source face, pick a target image or video, hit Start. Output lands in a folder named after the target.
2. **Live webcam** — pick a source face, hit Live, wait 10–30 seconds for the preview, then capture the window with OBS to feed it into a call or stream.

Source photo quality decides everything Front-facing, evenly lit, whole face visible, no sunglasses. A better source photo beats a better GPU.

It fails while installing requirements on Windows

Almost always the missing Visual Studio 2022 C++ build tools. Install them, then delete the venv and redo step 3.

ModuleNotFoundError: _tkinter on macOS

The GUI needs tkinter, which Homebrew ships separately: brew reinstall [email protected]

It starts but the swap never appears

Usually the models. Check that both .onnx files are in the models folder and are large files, not tiny HTML error pages.

gfpgan / basicsr errors

Known dependency clash. Reinstall those two from source: pip install git+https://github.com/xinntao/BasicSR.git@master, then pip uninstall gfpgan -y and pip install git+https://github.com/TencentARC/GFPGAN.git@master

Live mode is unusably slow

You're on CPU. Either set up CUDA (step 5) or accept a low frame rate — real-time swapping is genuinely heavy.

Why does it refuse some images?

The project has a built-in filter that blocks nudity and graphic content. That's deliberate and not a bug.

Consent isn't optional here Use faces only with permission and respect applicable rights and platform rules. Clearly label altered media where it could mislead. Do not make non-consensual sexual or deceptive content. This guide is not a legal determination.

Repo: **[github.com/hacksider/Deep-Live-Cam](https://github.com/hacksider/Deep-Live-Cam)** More free tools and setup guides at [rexxulabs.com](https://rexxulabs.com/) — follow [@rexxu.ai](https://www.instagram.com/rexxu.ai/) for the next one.