derpfakes

This is a pipeline for generating images from videos to use as input to deepfakes. To do so, it has several stages of processing:

1. Generate frames from videos: produce full-frame images at regular intervals from input videos
2. Isolate faces in video frames: identify and crop faces from every full-frame image
3. Recognize faces in isolated face images: recognize the face in each cropped image
4. Categorize images: categorize each cropped face based on who is in it

Installation Requirements

• Python3
• ffmpeg
• autocrop
• face_recognition

Usage

python run.py [args]

Arguments

• --test: Don't actually run any processor-intensive jobs (e.g. ffmpeg)
• --clearall: Clear work and output folders before doing any work
• --clear: Clear the working folder of the specified pipeline segment
• --verbose: Show lots of logs

Preparation

The pipeline requires source material in the two subfolders within the input folder.

• known_faces: put images of people you want to identify here, and name each file according to the categorization you want later. e.g. Albert Einstein.jpg and Emmy Noether.jpg.
• source_videos: and in here, put videos that contain the people you want identified

Running

When derpfakes starts, it creates a work folder in which temporary files are stored between pipeline stages. Because each stage might be a long-running process that might crash, the files stored in the work folder prevent work from being duplicated when derpfakes is run again.

If you want a fresh start, don't be shy to remove the work folder. You can also run python run.py --clearall which will destroy the work folder and then run the pipeline. If a particular stage seems broken, you can also delete the subfolder with the name of that stage, or run python run.py --clear [stage]. This is also a useful way to run the pipeline only after a certain stage.