search

shaoanlu / faceswap-GAN

A denoising autoencoder + adversarial losses and attention mechanisms for face swapping.
3.39k stars 841 forks source link

I trained 60,000 times, but the final model is very bad. Does anyone know what the reason is? #121

Open horizonheart opened 5 years ago

MoneyOfYang commented 5 years ago

我也是，莫名...

steveseguin commented 5 years ago

I've trained to around 40,000 and then to over 100,000 at 256x256; the final result is sorta okay, but it has an odd checkerboard pattern in most results and is softer than the input by a fair margin. The results at 64x64 hid these issues, but are pretty obvious at the higher resolution. I want to say that it feels like the original face-swap code without the adversarial component did better, so long as you aligned features. Certain features, like teeth, seem to benefit from the discriminator, but don't render as well as one might hope still.

dpk01singh commented 4 years ago

@steveseguin is this model yield good result on more number of images like 100000 images or so ?

steveseguin commented 4 years ago

If i recall, 1000 images of the same person got decent results. I can’t say what the ideal is though, as I was using around 80,000 images in my tests.

This was all quite some time ago, so I’m just spitballing.

On Sun, Mar 8, 2020 at 7:49 PM dpk Singh notifications@github.com wrote:

@steveseguin https://github.com/steveseguin is this model yield good result on more number of images like 100000 images or so ?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/shaoanlu/faceswap-GAN/issues/121?email_source=notifications&email_token=AATU2UUXAVENUWSSUWG55YDRGQVJJA5CNFSM4GOKXY52YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEOFFEHY#issuecomment-596267551, or unsubscribe https://github.com/notifications/unsubscribe-auth/AATU2UXEVBNLDKZ4IMI3RLTRGQVJJANCNFSM4GOKXY5Q .

--

Steve Seguin

steveseguin commented 4 years ago

I got good results with a custom data set that was properly aligned, decent resolution/sharpness, and about 3 days on a Titan RTX.

Mixing faces of different people reduced the quality when training. I used tens of thousands of faces, typically.

I'm not actually sure which Faceswap repo I was using at this point. I tried a dozen different ones and didn't find any all that great.

Hope that helps.

On Sun., Mar. 8, 2020, 7:49 p.m. dpk Singh, notifications@github.com wrote:

@steveseguin https://github.com/steveseguin is this model yield good result on more number of images like 100000 images or so ?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/shaoanlu/faceswap-GAN/issues/121?email_source=notifications&email_token=AATU2UUXAVENUWSSUWG55YDRGQVJJA5CNFSM4GOKXY52YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEOFFEHY#issuecomment-596267551, or unsubscribe https://github.com/notifications/unsubscribe-auth/AATU2UXEVBNLDKZ4IMI3RLTRGQVJJANCNFSM4GOKXY5Q .