▶️ Demo

▶️ AI-generated Elon Musk

Avatars for Skype and Zoom. Democratized.

Disclaimer: This project is unrelated to Samsung AI Center.

• conda
• CUDA

• Linux (with 1080 Ti GPU): 33 fps
• Mac OSX (MacBook Pro 2018; no GPU): ~1 fps

1. Download model's weights from Google Drive or Yandex.Disk [716 MB]
2. Place vox-adv-cpk.pth.tar file in the root directory

Linux uses v4l2loopback to create virtual camera.

Install avatarify dependencies (sudo privelege is required):

bash scripts/install.sh

For Mac it's quite difficult to create a virtual camera, so we'll use CamTwist app:

1. Download and install CamTwist from here. It's easy.
2. Setup avatarify conda environment with all required dependencies:

source scripts/install_mac.sh

Coming soon!

Copy your avatars into avatars folder. Crop pictures to make them square. Prefer pictures with uniform background.

Your web cam must be plugged-in. You can choose your camera by changing CAMID in scripts/settings.sh script.

Run Skype or Zoom only when Avatarify is started.

It is supposed that there is only one web cam connected to the computer at /dev/video0. The run script will create virtual camera /dev/video9. You can change these settings in scripts/settings.sh.

You can use command v4l2-ctl --list-devices to list all devices in your system. For example, if the web camera is /dev/video1 then the device id is 1.

Run:

bash run.sh

1. Run:

bash run_mac.sh

2. Go to CamTwist.
3. Choose Desktop+ and press Select.
4. In the Settings section choose Confine to Application Window and select python (avatarify) from the drop-down menu.

It is recommended to use the avatar overlay and the zoom in/out function to align your face in the preview window as closely as possible in proportion and position to the target avatar. When you have aligned, hit 'X' to use this frame as reference to drive the rest of the animation.

Alternatively, you can hit 'F' for the software to attempt to find a better reference frame itself. This will slow down the framerate, but while this is happening, you can keep moving your head around: the preview window will flash green when it finds your facial pose is a closer match to the avatar than the one it is currently using. You will see two numbers displayed as well: the first number is how closely you are currently aligned to the avatar, and the second number is how closely the reference frame is aligned.

You want to get the first number as small as possible - around 10 is usually a good alignment. When you are done, press 'F' again to exit reference frame search mode.

You don't need to be exact, and some other configurations can yield better results still, but it's usually a good starting point.

Go to Settings -> Audio & Video, choose avatarify or CamTwist camera.

Go to Settings -> Video and choose avatarify or CamTwist from Camera drop-down menu.

Make a call, allow browser using cameras, click on Settings icon, choose avatarify in Video settings drop-down menu.

Our goal is to democratize deepfake avatars. To make the technology even more accessible, we have to tackle two major problems:

1. Add support for more platforms (Linux and Mac are already supported).
2. Optimize neural network run-time. Running network real-time on CPU is of high priority.

Please make pull requests if you have any improvements or bug-fixes.

• Zoom/Skype doesn't see avatarify camera. Restart Zoom/Skype and try again.
• Avatar image is frozen: In Zoom, try Stop and Start Video.
• bash run_mac.sh crashes with "Cannot open camera": Try to change CAMID in run_mac.sh from 0 to 1 or 2
• pipe:0: Invalid data found when processing input: Make sure CAMID in scripts/settings.sh is correct. Use v4l2-ctl --list-devices to query available devices.
• ASSERT: "false" in file qasciikey.cpp, line 501. If you have several keyboard layouts, switch to English layout.

• Avatrify uses First Order Motion Model for generating avatars.