FaceCraft FakeApp

FakeApp Makes Faceswaps Easy

So how does this great algorithm works?

We can break it down into 3 major steps:

HOW IT WORKS

1. First step - Collect pictures/videos about your to-be-faked target person and selecting a video to be modified

2. Second step - Align

We only want to swap faces so our first step must be to identify faces on pictures. Once we found their position we can also identify their orientation and size. Using this information we can warp all picture to a general format.

Face detection is a traditional machine learning problem. One common way to do that is to use the Histogram of Oriented Gradients (or simply  HOG). This method looks at each pixel of our picture and compares its darkness to the surrounding pixels. Then we can draw an arrow (meaning  gradient) from the lighter to the darker points. We will use this representation of our original images. There is a HOG pattern that was trained on thousands of faces. We can compare it to our converted images to find areas that are very similar thereby identifying faces.

3. Third step - Train

Using our extracted face pictures we can train an autoencoder. Autoencoder is a type of neural network (the computational system behind deep learning). Its input is a picture and its output is the same picture. Doesn’t make any sense, right? 

Autoencoders consist of 2 parts: an encoder and a decoder. Encoder learns a much shorter representation of the input data while the decoder is able to transform it back to the original data. So encoder compress the data into a shorter size and decoder translate is back to the original longer format.

Training Preview

We train our model on our target pictures so our network will learn only how to encode a short representation into our target persons’s face. It means even if our input will be a different face our model will convert it into the original person’s face.

Now we can identify faces in our target video (video is just a bunch of pictures) then run them through our trained model converting the original actor to the target person.

4. Fourth step - Merge

In this phase we will take the new pictures created by the autoencoder. In the alignment phase we saved their original orientation and size so doing some backward engineering we can put them back into the original picture. Out of the modified pictures we can recreate our video.

And that’s all, our new deepfake video is ready!

Follow Us On Instagram