Open QLaHPD opened 3 years ago
Yes, that's what I was talking about.
@QLaHPD Thanks for your suggestion.
You have proposed three ways to improve.
It would be very great if you have resources to have a try~
Ok, I can try to code this in a fork soon, and train in Colab.
Ok, I can try to code this in a fork soon, and train in Colab.
how are they going
I think it would be very useful to add more discriminators, from the tests I have done with conditional GANs, it seems that having several discriminators with different levels of reception fields increases the support of the distributions as well as the stability and quality of the images (maybe can remove the artifacts like the ones on Figure 11 in the paper). It would also be interesting to try a discriminator with MLP Mixer architecture (https://github.com/jaketae/mlp-mixer, https://github.com/sradc/patchless_mlp_mixer) since the paper shows that the "way" that this type of architecture selects the features is different from what a CNN does, so maybe it helps the Generator to not create certain types of artifacts.
Also, I'm not sure, but does the ESRGAN architecture have multiple noise inputs? If not, I also think it would be useful to add noise to each res-block, since more noise usually helps.