interdigitalinc / featurestyleencoder Goto Github PK

Hi, It great work

Since there are no boundaries provided in the repo, I've tried my own to change facial attributes. Unfortunately, my boundaries work well if I use them only with pSp or e4e, but most of them do not work well with your encoder. They produce very ugly output images. I wonder why, since you also use StyleGAN2. So what is the reason and the sollution? And how do you look for directions using InterFaceGAN if its originally used for StyleGAN1?

The training code in trainer.py line 180 (features = [None]*k + [fea] + [None]*(17-k))should be modified as:

if training_mode:
    features = None
else:
    features = [None]*k + [fea] + [None]*(17-k)

I have received an email that identifies this problem and proposes a modification. This modifications has no impact at test time but do impact the training. Without this the model is not trained on two inversions.

Hi, thanks for contributing this excellent work.

Could you please tell whether the L2 recon. loss is applied on \hat{x}_2 = G(w, F) for the real image case? I feel a bit confused about this point when reading the paper. Thanks in advance.

Hi Xu-Yao,
I've tried running the code and found the code under the pixel2style2pixel/ folder is missing
I've tried using the code from the original implementation instead, but there are at least several changes like Generator's forward method getting input features and possibly many additional changes.

I'm not sure if you are still in the process of updating the repo, but thought I'd remind you just in case you simply forgot to add your adapted pixel2style2pixel/ code.

Looking forward to testing your encoder. looks very promising

Hi,
Where can we found boundary files to edit images in latent Space ?

Hi, It great work

Hello,
Thank you for your valuable work. I have recently implemented In-Domain Gan inversion for editing attributes of binary images. As it didn't worked well on external real image editing I consider your methods which seems has met some of the previous limitations.
My question is which version of StyleGAN is better for this method? I have trained StyleGAN2-ADA on my images. I wanted to know if the augmentation method in this version affects the inversion or editing using your approach.
What do you recommend?

Thank you in advance for your help.

Running the test.py file shows that there is no alex.pth

When I run the train.py file, setting the batchsize to 2 will increase the running memory by more than two times. What's going on?

Love the paper
Any chance the code will be released soon?

Hi, sorry for bothering again.

I wonder if it is convenient for you to share the structure of the network? Specifically, the feature prediction branch and latent prediction branch. Thanks.

Running the test.py file shows that there is no alex.pth