can I use PatchSWD for AutoEncoder reconstruction?

[iperov]

Thanks for your work.

can I use this loss for AutoEncoder reconstruction?

Currently I am using multiple patch sizes in order to get more GAN-like patterns of reconstruction.

Are there any additional approaches to get better quality?

[ariel415el]

Sure you can. I tried this too but it does not seem to be comparable to GAN

The results looks somewhat similar to MMD++ here https://github.com/ariel415el/PerceptualLossExperiments#5-generative-models

If interested you can read this paper that analyzes patch distribution losses and tests them in super-rseolution setting