the effect of pretrained model about gpen

[yangxy]

In our experiments, if the degradation is not that severe, no pre-trained model can achieve comparable results, which is somewhat not consistent with the idea claimed in our paper. However, in cases where the face is severely degraded such as 64x FSR, no pre-trained model can hardly produce any clear results, while GPEN still works well.

Hope this helps.

[yangxy]

In our experiments, if the degradation is not that severe, no pre-trained model can achieve comparable results, which is somewhat not consistent with the idea claimed in our paper. However, in cases where the face is severely degraded such as 64x FSR, no pre-trained model can hardly produce any clear results, while GPEN still works well.
Hope this helps.

Thanks for answering.
noise add vs noise cat, should I retrain a stylegan2 model with noise cat?

Yes if you want to reproduce our results. Actually, the noise-add version of GPEN can work quite well too, in such case you can simply adopt a well-trained StyleGAN.

Thanks
Lf is similar to the perceptual loss [19] but it is based on the discriminator rather than the pre-trained VGG network to fit our task.
If I use pre-trained VGG , the effect will be similar?

To be honest, a single L1 loss leads to quite nice results.

[lxy2017]

In our experiments, if the degradation is not that severe, no pre-trained model can achieve comparable results, which is somewhat not consistent with the idea claimed in our paper. However, in cases where the face is severely degraded such as 64x FSR, no pre-trained model can hardly produce any clear results, while GPEN still works well.

Hope this helps.

Thanks for answering.
noise add vs noise cat, should I retrain a stylegan2 model with noise cat?

[yangxy]

In our experiments, if the degradation is not that severe, no pre-trained model can achieve comparable results, which is somewhat not consistent with the idea claimed in our paper. However, in cases where the face is severely degraded such as 64x FSR, no pre-trained model can hardly produce any clear results, while GPEN still works well.
Hope this helps.

Thanks for answering.
noise add vs noise cat, should I retrain a stylegan2 model with noise cat?

Yes if you want to reproduce our results. Actually, the noise-add version of GPEN can work quite well too, in such case you can simply adopt a well-trained StyleGAN.

[lxy2017]

In our experiments, if the degradation is not that severe, no pre-trained model can achieve comparable results, which is somewhat not consistent with the idea claimed in our paper. However, in cases where the face is severely degraded such as 64x FSR, no pre-trained model can hardly produce any clear results, while GPEN still works well.
Hope this helps.

Thanks for answering.
noise add vs noise cat, should I retrain a stylegan2 model with noise cat?

Yes if you want to reproduce our results. Actually, the noise-add version of GPEN can work quite well too, in such case you can simply adopt a well-trained StyleGAN.

Thanks
Lf is similar to the perceptual loss [19] but it is based on the discriminator rather than the pre-trained VGG network to fit our task.
If I use pre-trained VGG , the effect will be similar?

[yangxy]

In our experiments, if the degradation is not that severe, no pre-trained model can achieve comparable results, which is somewhat not consistent with the idea claimed in our paper. However, in cases where the face is severely degraded such as 64x FSR, no pre-trained model can hardly produce any clear results, while GPEN still works well.
Hope this helps.

Thanks for answering.
noise add vs noise cat, should I retrain a stylegan2 model with noise cat?

Yes if you want to reproduce our results. Actually, the noise-add version of GPEN can work quite well too, in such case you can simply adopt a well-trained StyleGAN.

Thanks
Lf is similar to the perceptual loss [19] but it is based on the discriminator rather than the pre-trained VGG network to fit our task.
If I use pre-trained VGG , the effect will be similar?

To be honest, a single L1 loss leads to quite nice results.

plus LA of course

[lxy2017]

In our experiments, if the degradation is not that severe, no pre-trained model can achieve comparable results, which is somewhat not consistent with the idea claimed in our paper. However, in cases where the face is severely degraded such as 64x FSR, no pre-trained model can hardly produce any clear results, while GPEN still works well.
Hope this helps.

Thanks for answering.
noise add vs noise cat, should I retrain a stylegan2 model with noise cat?

Yes if you want to reproduce our results. Actually, the noise-add version of GPEN can work quite well too, in such case you can simply adopt a well-trained StyleGAN.

Thanks
Lf is similar to the perceptual loss [19] but it is based on the discriminator rather than the pre-trained VGG network to fit our task.
If I use pre-trained VGG , the effect will be similar?

To be honest, a single L1 loss leads to quite nice results.

plus LA of course

Thanks!
If you use a pre-trained model for discriminator?