Open EXJUSTICE opened 5 years ago
For sure! This implementation was only for learning purpose, a simple mask should improve a lot the results, but also complicates the code, so I prefer to let it this way. If you need better results your link is a better place to start maybe.
Interesting, can you elaborate on what do you mean by mask here? Facial landmarks?
I will fork your code and modify it with a more complex model and observe, would make for an interesting beginners tutorial!
Not facial landmarks, the first way to improve this, IMO, should be create an R-CNN Mask or something similar in order to use more the original pixel or the inferenced pixel depending on this mask. The link you show me use this technique to improve the quality among others
Very nice implementation, I'll look into tweaking it. However, Do you think that the poor performance is due to the simplicity of the architecture?
Taking inspiration from the below link could help? https://github.com/shaoanlu/faceswap-GAN/blob/master/README.md