Bad inversion quality for custom image.

[picksh]

Hi,
Thank you for releasing the data alignment and real-image inversion scripts. I followed the instruction in README to first align the image and then invert real image with PTI using the PTI weights that you offered. However, I found that the Face was very different after inverting. The result is shown below:

And I also try to ajust the LPIPS_value_threshold to be lower, but the result is still not satisfying. I also put the result below:

I think the inversion result in your paper is much better so do you have some advice or did I just ignore some steps?
The image I used is from DeepFashion-MultiModal dataset.
Thank you for your help.

[stylegan-human]

Hi,
I cloned a fresh repo and followed the instruction to invert the same image. (The aligned image and the inverted result are listed below.) The face I got is different than yours. It looks like your alignment image is squeezed horizontally.

[picksh]

Thanks for your reply. I think there might be some mistake with my alignment and I will try it again. But I think the inversion result that you give above is still different from the original image(the face). So is it a normal result when you work with your project?( May be a limitation of StyleGAN inversion itself..)

[c1a1o1]

Can I see your inversion model?
68545849@qq.com

[ds-nat-j]

Thanks for your reply. I think there might be some mistake with my alignment and I will try it again. But I think the inversion result that you give above is still different from the original image(the face). So is it a normal result when you work with your project?( May be a limitation of StyleGAN inversion itself..)

I've seen that the GAN inversion for Full body images usually occurs outside the editable regions of the StyleGAN Latent space. So to counter this, there is some trade-off on reconstruction quality and I think the authors of PTI went ahead with it anyway, which is why the reconstruction of these images look pretty bad.

To get the best editable results I think the HyperStyle approach that promises "to modulate StyleGAN's weights to faithfully express a given image in editable regions of the latent space" is something the authors of the repository should consider exploring.

cc: @stylegan-human