landmark customization related question

[mithrandir]

hi, i am new to faceswap. i would like to know if it is possible to customize the landmarks in a way that i am reducing the faceswap area to the eyes only? i would like to swap only the eyes between 2 videos, if it is possible.

[torzdf]

Sure, it's possible. That's going to require some coding on your part though, as our whole codebase is built to work on faces.

Ultimately a NN doesn't know what it's learning, just that it learns. If you want to give it eyes, it can learn eyes.

[mithrandir]

thanks for the reply. it sounds interesting. actually i just started to learn machine learning. i think i understand the basic concept of dl. i have some experiences with python by creating various scripts for api calls, web scraping and other office automation purposes, but i don't have the developer background, so i am not sure where to start if i want to make this change.

i know that the alignments files were previously json files, ones i had a look on a file, if i am right it is just a collection of the landmark points. i guess that the fsa files serve the same purpose, so maybe i need to make 2 modifications:

• modify how the code is collecting these landmarks when creating the alignments files
• and then modify how the dl algorithm is using the landmarks
  i guess only changing the alignments files is not enough, right? the dl would get an error if the alingments file has less landmarks?

[mithrandir]

in the meanwhile i just had a look on extract.py, train.py, convert.py, also found a couple of videos on youtube, which expain how faceswap works in python. so maybe if i will check all these sources i will be able handle this case. however it might happen that later i will come back with some more concrete questions.