Closed c1a1o1 closed 5 years ago
First, I'm not the author of the paper. I'm doing research in face editing so I implemented AttGAN in PyTorch.
To be honest, the model does not perform well on all images. As you can see in my custom results, some faces look not good.
The glasses are not quite visible in your images. My suggestion is that you can try with a larger test value rather than just 1. For example, test with --test_int_max 1.5
, you will probably get better glasses. On the other hand, you mentioned that you'd like to generate sunglasses. The pretrained model and the implemented code does not support it because the labels in CelebA are like with glasses / without glasses
. This code generates a face with desired attributes, but we are not able to constraint the style of each attribute. If you want to control the style, you have to implement the Mutual Information Maximization as the paper describes. I have no plan to do it so far.
Thank you very much! Maybe you can look at this work for face editing! https://github.com/quolc/neural-collage
This work is pretty cool! I'll read it. Thank you for the suggestion!
I hope you can use pytorch to reproduce it, thank you.
When we tested with test_slide, we couldn't get the glasses results shown in Figure 7 in your paper.
Can you give some advice?
It can be seen that it is impossible to generate dark glasses