joshua-wu/deepfakes_faceswap

has obvious signs of similar ps

zaxs002 opened this issue · 7 comments

训练之后测试感觉有明显的类似ps的痕迹.
看deepfakes上传的图片感觉几乎没有痕迹.
是要继续训练还是有其他的方法?
有没有人一起讨论一下?

Sorry guys..
In English.
Translated by google.

After training the test has obvious signs of similar ps.
Look at the image of deepfakes upload almost no trace.
Is it necessary to continue training or are there other ways?
Anyone discuss it?

I was training Cage-Trump models all night on Tesla K80, and the results were like this:
9f6ba088a87ba047163b99c3384e3f8d
c1abcde4a2e549ee59901495af66d7ef

Deepfakes reply me that:
7ec0b56f54b202a8992d85c029c4f774

I have seen serveral uploaded pictures.
Look like it's handled very well..But also high-definition.
Don't know if threre is a good idea?
Discuss everyone.

Clorr commented

In its latest post deepfakes proposes to use dlib to better align faces. This will improve the result.

Otherwise it is also due to the low-res of the training / generation. Have you tried higher resolution ?

Check this OpenCV SeamlessClone code by depfakacc:
https://gist.github.com/anonymous/8f468972d6286f403318c94bc6dbe382
I trained Trump-Putin model, and used 1adrianb/face-alignment, the result was better
https://gfycat.com/WebbedDirectGossamerwingedbutterfly

Clorr commented

Indeed this is nice. I want to integrate this in the main project, but Christmas did not let me much time...

Look at #16 , and don't hesitate to push code if you have some...

It gets better with more training.

@zaxs002 @junglewreck @HorselessHorseperson Hi ,i also have the same result,how do you solve it ?Continue training or train with more other pictures?
image