How to produce a deepfake

If your authoritarian governement controls all the media and floods the infromational space with fake news 24/7, you might be interested to learn how to make your own political message visible. Take the face of the person who took your political rights to reclaim them.

1. Deepfake is a method of video production that allows you fake footage of targeted personages by means of deep learning algorithms. In such videos, a person’s voice and behavior are simulated in recordings that do not correspond to real footage.

2. You need to set up a political and ethical framework, and note practical considerations:

  • What are the political implications?

  • Are you perpetuating power dynamics that inflict harm on oppressed groups?(Fake porn is one of multiple examples of how one should not use deepfake technology).

  • Punch up! For this manual, I will use the example of making a politician give an unrealistic speech.

3. The more realistic the deepfake is, the more complicated its production.For both sound and video you will need a machine with a powerful GPU.If you don’t have a suitable PC, you can use google colab - .

4. Start with the audio production. Do you need the selected character to speak? If not,ou can still achieve a lot of impact by making a politician appear in an unusual setting, which would save you a lot of effort. If you decide to only work with video, skip to step 7.

5. Work in English. For the moment, finding a neural network trained in other languages is challenging.

6. In order to make a voiceover for deepfake, use text-to-speech models like deepvoice3. First, you will need to do “speaker adaptation.” Training the model from scratch is possible, but more demanding.

7. Experiment with Facelab, the most used deepfake software. Facelab exists as a google colab and a github repository.

8.  Record footage with a model. You will later alter this because Facelab only replaces the face from your footage with the face of your target. Therefore, to produce a realistic deepfake the footage should be recorded in a realistic setting or else it requires postproduction to place the deepfaked face back on the body of the target.

9. Facelab has a lot of tutorials online but none of them explain postproduction. Postproduction can be done by replacing the background if the shooting of the model was done on a green screen. Otherwise, it is possible to replace the face with motion tracking in AfterEffects.