Scriptified version of v2.1. #92
Conversation
…tant progress and automatically restarts from last checkpoint. Previews are saved as png files. Losses are plotted using matplotlib.
|
Hi, |
|
Really appreciate slundell spending time making this PR, and sorry for postponing it for so long despite I'm the person asking for help. I got some feedback making me feel that I should remain this project less accessible for people without much coding experience. Obviously deepfakes is still a controversial topic, and I would like to keep this project from being closely associated with it. (Edit: I was trying to refer to the concept of deepfakes, but I chose the wrong wording that seems offensive to the deepfakes community.) As for now, I prefer to make this project deep-learning-oriented (as well as messy and loosely structured) instead of being a hands-on face-swapping tool. |
|
@shaoanlu Get off your high horse. You even show your work using some women's videos. |
|
People always will use the tools for good and bad, it's doesn't matter what kind of tool it is. |
Includes schedule copied from v2. Persitant progress and automatically restarts from last checkpoint. Uses prefetch_generator to increase GPU utilization. Previews are saved as png files. Losses are plotted using matplotlib. After 40000 iterations videos are automatically converted. This script requires a specific folder structure:
from-person/extracted-faces (folder with extracted faces of "face-donor")
to-person/extracted-faces (folder with extracted faces of "body-donor")
to-person/to-swap/ (video files to be converted)
These folders will automatically be created:
to-person/models (weights files, progress file, loss plot)
to-person/models/preview/ (images of conversions)
to-person/swapped/ (converted videos with new face)
Invoke like this:
python3 make-fake.py --source-dir ../data/cage --target-dir ../data/trump
There is a lot of cleanup to be done in this script, but it should be functional as-is.