Swapping more than face

Faces are kind of a consistent feature across the board and as far as current tech goes, easy to identify. Expanding that to the chin and jaw would probably be possible too, if we were to expand the area that we extract, train and convert.

However, with other bodyparts it depends on what you're actually training the model for. For example train a model with Trump's hands and a child's hands, you might actually be able to replace hands. (not that you would be able to tell the difference with Trump's tiny hands 😆 ) As for parts like shoulders, that would probably be a lot more difficult.

However, Hair on the head would probably be much more difficult as it is can take so many shapes, unlike a face which has the same landmarks every time. You'd probably have to get a model to recognize specific kinds of hair. Short brunette vs long blonde. Long hair poses a challenge also, because it covers more area. If you have to replace long hair with short hair, you have a problem that the model needs to fill in that area. Sometimes it might need to also suddenly add ears to a head. Major challenges!

Either way, the best way to find out if any of this works, is by gathering training data and trying to replace random stuff with it. These algorithms can be trained to recognize faces, but also other shapes like watches, coffee cups and bananas. I suggest people start trying to get these libraries to recognise and swap other things than faces.

[Comment edited for content]

deepfakes / faceswap-model

Swapping more than face #6