search

shaoanlu / faceswap-GAN

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

Documentation for high quality results parameters #147

Open vmartinv opened 5 years ago

vmartinv commented 5 years ago

Is there any place I can read about how to produce high-quality results? The pictures from the README look very well produced whereas the result I get from the collab demo is not as good. Is there any documentation about the parameters for that Trump-Cage example? Thank you so much

ghost commented 5 years ago

@vmartinv have you looked at https://github.com/shaoanlu/fewshot-face-translation-GAN dont have to train on face pairs with it

vmartinv commented 5 years ago

I prefer the Trump-Cage results, as I want some to use it as a benchmark for my thesis. This is the reason I would like to know more about the training process. Did you obtain the Trump-Cage results by just following the usage section in the readme? Thank you!

ghost commented 5 years ago

@vmartinv those were trained on two face pairs so the results would be better as the model would not work on unseen subjects like the few shot face translation GAN which I think you should use as a benchmark as recent works like https://nirkin.com/fsgan/ work with a single image without training face pairs. I have not tried training trump cage with this repo but have trained other pairs with not much success unless you spend more time training.

zhanglonghao1992 commented 5 years ago

@vmartinv I have been experimenting with Trump-Cage. Training data has a big impact on the results. When I add some images taken from the video, the results are not good. When I only use the Trump-Cage dataset form deepfakes' project, the results are much better. However, due to the limited number of images in the dataset, many extreme facial expressions and side faces results are bad. cage10

There are a lot of things in this project that need to be adjusted, such as the weight of multiple losses, the hyper-parameters in the phased training strategy. When it comes to video face-swapping, you need to test different kf_noise_coef and color correction methods. cage_swap_trump