How to make deepfake video clips for training

Step-by-step instructions for creating example real and deepfaked video clips for the ‘Spot the Fake’ activity, using both Windows and Linux — with free tools or browser-based generators.

This guide walks you through how to create short real and deepfaked video clips for training activities in shelters or workshops. These clips are meant for educational use only — not for social media pranks, political mayhem, or triggering your neighbour.

You can do this on:

  • A Windows machine, using free downloadable tools or browser-based services
  • A Linux machine, with command-line or GUI options
  • Both online and offline workflows, depending on your threat model and available time

Key principles

  • Keep clips under 30 seconds
  • Use neutral, non-triggering content (e.g. “Hello, I’m X. Welcome to the shelter.”)
  • Always disclose to participants what is fake, and how it was made
  • Never use real survivor, staff, or client identities in fake videos
  • Store and delete responsibly

Real video recording

You will need:

  • A phone or webcam (even 720p is fine)
  • A quiet space with decent lighting
  • Someone willing to say a short script
  • No editing needed unless you want to trim the clip

Suggested free trimming tools:

  • Windows: Use the built-in Photos app → Edit → Trim
  • Linux: Use Shotcut, Kdenlive, or run:
ffmpeg -i input.mp4 -ss 00:00:01 -to 00:00:29 -c copy trimmed.mp4

Creating deepfake versions

You can either use online AI tools (fast, no install) or offline tools (more control, privacy).

Windows: Using online tools

Easiest for most users, only good if you do not mind uploading footage to the cloud.

Option 1: DeepBrain AI Studios

  • GoTo DeepBrain AI Studios
  • Create a free account (limited usage)
  • Upload a script and pick a face (or create a clone)
  • Generate a clip — usually under a minute
  • Download the video (MP4)

Option 2: HeyGen

  • GoTo HeyGen
  • Choose a presenter or upload your own photo
  • Add a text script
  • Supports multiple languages and accents
  • Free trial with watermark

Option 3: Synthesia.io

  • GoTo Synthesia.io
  • Polished avatars, very slick UX
  • Requires account
  • Free trial includes a few videos

All online platforms store your clips. Use only generic content and consider using burner emails.

Windows: Using free offline tools

Option 1: Avatarify (open-source, real-time deepfake)

pip install avatarify
  • Use with a webcam + script reader
  • Overlay a celebrity or stock face in real time

Option 2: DeepFaceLab

  • Download from GitHub
  • Requires strong GPU and patience
  • Best for realism, but setup is advanced

Linux: Using free tools

Option 1: First-order Motion Model (FOMM)

git clone https://github.com/AliaksandrSiarohin/first-order-model
cd first-order-model
pip install -r requirements.txt

  • Feed it:

    • A still image (face)
    • A driving video (you or actor reading script)
  • Outputs animated video

Option 2: DeepFaceLive (Linux-native build)

There is a Linux-native build, although it is far less documented and requires a bit more fiddling with dependencies like dlib, onnxruntime, and specific versions of ffmpeg. But it works:

  • Clone the project: git clone https://github.com/iperov/DeepFaceLive.git

  • Follow the Linux setup instructions (not as polished, but workable):

    • Install dependencies like Python 3.8–3.10, onnxruntime, torch, opencv, and dlib.
    • Use virtualenv to keep things tidy.
    • Expect to troubleshoot a bit—especially with CUDA if you want GPU acceleration. It is rather system heavy.
  • Good for generating video output with voice + face overlay.
  • Less real-time smooth than on Windows, but usable with some patience.

What to prepare

  • 3 to 4 real clips with simple introductions
  • 3 to 4 fake clips, generated from same or similar scripts
  • Optional: One blended clip where only part is altered (e.g. voice vs face)

Use consistent lighting and tone so the difference is subtle — it makes the spotting game harder (and more fun).

Privacy and ethics

Do:

  • Get informed consent from real speakers
  • Use made-up names and harmless scripts
  • Explain how and why the fake was made

Do not:

  • Use videos of real survivors, children, or sensitive stories
  • Use deepfake generators without checking their terms
  • Forget to delete cached or leftover training data