search

deepfakes / faceswap-playground

User dedicated repo for the faceswap project
306 stars 194 forks source link

Understanding GAN model's choices #297

Closed Luna419 closed 5 years ago

Luna419 commented 5 years ago

I'm currently working with the GAN model and I have pretty good results. I'm now trying to understand the choices that were made for the training. For example, why is it interesting to calculate the loss of the discriminators (d_lossA and d_lossB) with the mixup option ?

deepfakes commented 5 years ago

GAN was removed from this repo a long time ago.

deyu260 commented 5 years ago

I'm currently working with the GAN model and I have pretty good results. I'm now trying to understand the choices that were made for the training. For example, why is it interesting to calculate the loss of the discriminators (d_lossA and d_lossB) with the mixup option ?

can you put your result from gan model? i am doing the same thing .

Luna419 commented 5 years ago

Here is a 2s video of deepfake (Oliver to Fallon). The conversion to gif format lowered the quality but you can see that the face is quite ok. However the edges of the part rebuilt by the gan can clearly be seen.

John_Oliver_Trim_2s_120000ep_GAN_blur