How To Do AI Video Deepfakes

Biohazard
Jul 18
5 min read

AI Video Deepfake Guide

Let's go deep on AI video deepfake — covering full pipelines from source material to real-time delivery. Here's a(n) article / guide on how to do AI video deepfakes.

Source Material Requirements

Quality	Face Images	Video	Source
Minimum	5-10 clean frontal shots	30 seconds of frontal video	LinkedIn, company website, YouTube
Good	20+ varied angles/lighting	2-5 minutes with expressions	Internal videos, interviews, conference talks
Ideal	50+ from multiple angles/emotions	10+ minutes with varied head movement, lighting	Professional headshots + video catalog

Extraction tools:

bash

# Extract frames from source video
ffmpeg -i source_video.mp4 -vf fps=1 frames/frame_%04d.png

# Extract face crops (using facedetection)
ffmpeg -i source_video.mp4 -vf "select=gt(scene\,0.4),crop=iw/2:ih/2:0:0" -vsync vfr faces/%04d.png

# Better: use a dedicated face extraction script
git clone https://github.com/deepfakes/faceswap.git
cd faceswap
python faceswap.py extract -i source_video.mp4 -o faces/

Tool Selection

Option A: DeepFaceLab (Highest Quality, Local, GPU-Intensive)

The industry standard for high-quality deepfakes. Requires NVIDIA GPU with 8GB+ VRAM.

bash

# Download from https://github.com/iperov/DeepFaceLab
# Extract and run with GUI or batch scripts

# Typical workflow (via batch scripts):
# 1. Extract video frames
2) extract images from video data_src.bat

# 2. Extract faces from source
3) extract faces from images data_src.bat

# 3. Extract target video frames
4) extract images from video data_dst.bat

# 4. Extract target faces
5) extract faces from images data_dst.bat

# 5. Train (runs for hours/days depending on quality)
6) train Quick96.bat

# 6. Convert (apply trained model)
7) convert Quick96.bat

# 7. Merge with original video
8) merged to mp4.bat

Quality vs Time tradeoff:

Model	Training Time	Quality	VRAM
Quick96	2-8 hours	Low-medium	4GB
SAEHD	12-72 hours	Medium-high	6GB
AMP	24-96+ hours	Very high	8GB+

Option B: FaceFusion (Easier Setup, Good Quality)

bash

git clone https://github.com/facefusion/facefusion.git
cd facefusion
python install.py
python run.py

# Opens a web UI at http://localhost:8000
# Upload source face image, target video, click Process

For command-line automation:

bash

python run.py --source face.png --target video.mp4 --output result.mp4 --execution-providers cuda

Option C: Roop / Roop-Unleashed (Simplest, Single-Click)

bash

git clone https://github.com/s0md3v/roop.git
cd roop
pip install -r requirements.txt
python run.py

# Or Roop-Unleashed with more features:
git clone https://github.com/C0untFloyd/roop-unleashed.git
cd roop-unleashed
python installer.py
python run.py

Option D: ComfyUI (Most Flexible, Advanced Workflows)

bash

git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
pip install -r requirements.txt
python main.py

# Install custom nodes for face swapping:
# - ComfyUI-Impact-Pack (includes face swap)
# - ComfyUI-FaceSwap (using InsightFace)
# - Reactor (fast face swap node)

Real-Time Video Deepfake (Live Calls)

Deep-Live-Cam

Best for real-time face swapping in video calls (Zoom, Teams, Meet, etc.).

bash

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

# Install dependencies
python install.py

# Run
python run.py --execution-provider cuda

# Requires:
# - Source face image
# - Webcam input
# - OBS Virtual Camera or direct webcam output

OBS Virtual Camera setup for use in any video call app:

bash

# In OBS:
1. Add Video Capture Device (your real webcam)
2. Add a filter → Apply face swap from Deep-Live-Cam
3. Tools → VirtualCam → Start
4. In Zoom/Teams/Meet → Select "OBS Virtual Camera" as your camera

Real-time avatar with audio sync

bash

# Combine with voice cloning for full impersonation
# Use Deep-Live-Cam for face + ElevenLabs/Coqui for voice
# Deliver via OBS Virtual Camera + VB-Cable Virtual Audio Cable

Delivery Vectors

Pre-Recorded Video Phishing

bash

# Create a fake CEO announcement video
# 1. Find a real CEO video (town hall, earnings call)
# 2. Deepfake the face
# 3. Voice clone the audio
# 4. Sync and export

ffmpeg -i deepfake_video.mp4 -i cloned_audio.wav -c:v copy -c:a aac final_output.mp4

# Host on a legitimate-looking URL or email attachment
# Test if the organization's security awareness training catches it

Live Video Call Impersonation

Setup:

Hardware: Webcam + decent lighting that matches the target's usual environment
Software: Deep-Live-Cam → OBS Virtual Camera
Audio: Voicemod or VB-Cable + ElevenLabs real-time voice cloning
Platform: Zoom/Teams/Meet — join as the executive

Pretext:

"I'm calling from the airport — my video is glitchy but I need to discuss something urgent before the board meeting."

The glitchy video excuse covers any artifacts or frame drops from the deepfake.

Phishing Campaign with Deepfake Video

bash

# Embed a deepfake video thumbnail in a phishing email
# When clicked, plays the video that appears to be the CEO
# asking the employee to "click the link below to confirm"

# Alternative: use as a landing page component
# - Clone the company portal
# - Add a "Message from CEO" video (deepfake)
# - Video says "Enter credentials to view secure message"

Testing Liveness Detection

If the organization uses liveness/anti-spoofing on video calls:

Liveness Check	Bypass Method
Head movement challenge ("Turn left")	Use a video source with pre-recorded head movements, or Deep-Live-Cam tracks your real head movement
Eye blink detection	Deep-Live-Cam and Roop both support real-time eye blink transfer
Mouth movement / lip sync	Real-time tools sync lip movements to your actual speech
Background inconsistency	Use a solid background or a green screen matching the target's usual environment
Lighting mismatch	Match your room lighting to what the target would have. Add a fake window light, desk lamp
Reflection check	Wear glasses if the target does. Some liveness tools check for screen reflection on lenses

Full Attack Pipeline (Ready-To-Use)

bash

# Phase 1: Gather source material
yt-dlp -f bestvideo+bestaudio "https://youtube.com/watch?v=CEO_INTERVIEW" -o ceo_source.mp4
ffmpeg -i ceo_source.mp4 -vf fps=1/5 frames/ceo_%04d.png

# Phase 2: Extract face set
python faceswap.py extract -i ceo_source.mp4 -o ceo_faces/

# Phase 3: Train (DeepFaceLab - SAEHD)
# This runs for 24-48 hours on a decent GPU
# Results go to workspace/model/

# Phase 4: Record target video script
# Record yourself saying: "I need you to process payroll early this week..."
ffmpeg -f v4l2 -i /dev/video0 -f alsa -i hw:0 target_recording.mp4

# Phase 5: Apply deepfake
python run.py --source ceo_faces/00001.png --target target_recording.mp4 --output final_deepfake.mp4

# Phase 6: Voice clone (ElevenLabs)
python voice_clone.py --source ceo_audio.wav --text "script.txt" --output cloned_audio.wav

# Phase 7: Replace audio
ffmpeg -i final_deepfake.mp4 -i cloned_audio.wav -c:v copy -c:a aac -shortest delivery_ready.mp4

# Phase 8: Deliver
# - Upload to a phishing landing page
# - Attach to email (if <25MB)
# - Host on a file-sharing service

Detection & Reporting

Detection Method	What It Looks For	Likely to Catch
Human review	Uncanny valley, eye movement artifacts, lip sync errors	Medium — better approaches
Deepware Scanner	GAN fingerprints, blending artifacts	Low-to-medium (OBS w/ compression bypasses most)
Microsoft Video Authenticator	Edge blending, boundary artifacts	Low with high-quality models
Intel FakeCatcher	Heartbeat/blood flow detection	Medium — requires dedicated hardware
Sensity AI	Model-specific signatures	Varies by tool used
Exif/metadata analysis	Editor stamps, processing history	High if no cleanup

Report example:

markdown

## Finding V-002: Video Deepfake Susceptibility

Severity: Critical
Vector: CEO impersonation via fake town hall video posted to internal Slack
Tool Used: DeepFaceLab (SAEHD, 48hr training) + ElevenLabs

Result: 14/20 employees who watched the video believed it was real.
2 employees flagged it as suspicious. No one verified via out-of-band channels.

Remediation:
1. Implement liveness detection on video calls (challenge-response)
2. Establish out-of-band verification protocol for video-based requests
3. Deploy automated deepfake detection tools (Sensity, Deepware)
4. Conduct deepfake awareness training for all staff
5. Create a "verified channel" for executive communications