search

shaoanlu / faceswap-GAN

A denoising autoencoder + adversarial losses and attention mechanisms for face swapping.
3.37k stars 842 forks source link

Head swap #107

Open kulievvitaly opened 6 years ago

kulievvitaly commented 6 years ago

Thanks for the work you've done. In many cases, replacing the head (not the face) will look more realistic. I tried changing the dataset, but the program still selects only the faces. Is it possible to enlarge the area around the face in order to generate the whole head?

Yoba777 commented 6 years ago

Increase roi_coef or make it negative. It decides which offset is taken from original face bounding box

kulievvitaly commented 6 years ago

I tried to change this parameter. The problem is that the initial window with which the neural network works is much smaller than the size of the head.

Yoba777 commented 6 years ago

Если сделать его равным -1, окно увеличится в два раза - голова вполне может поместиться)

7 июля 2018 г., 2:47 пользователь kulievvitaly notifications@github.com написал:

Я пробовал изменять этот параметр. Проблема в том, что начальное окно с которым работает нейросеть существенно меньше размеров головы.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/shaoanlu/faceswap-GAN/issues/107#issuecomment-403129962, or mute the thread https://github.com/notifications/unsubscribe-auth/AlrUC4X_SixZsolut7I5xqre5_-L6rAvks5uD77sgaJpZM4VCZ8q .

RaineyWu commented 5 years ago

Thanks for the work you've done. In many cases, replacing the head (not the face) will look more realistic. I tried changing the dataset, but the program still selects only the faces. Is it possible to enlarge the area around the face in order to generate the whole head?

my dataset is just the face and the output is not so good with(5W iters) , have you tried change the dataset to head yet?

kalmsy commented 5 years ago

@Yoba777 Thank you is there any chance you can please can you post detailed instructions on how to head swap?

I think i can manually create dataset more zoomed out. But in my case to be honest i am using the fakeapp wrapper on GAN. @kulievvitaly do you think if we have zoomed out dataset and parameter changed it will work? Do you know how i can change roi_coef in fakeapp?

jackweiwang commented 5 years ago

I am sorry, I want change the roi_coef parameter, but It does not exist.