Can this method be trained on images with non-1:1 resolution.

[super782]

Thanks to open source for this wonderful work. I have a question if this method is only supporting 1:1 resolution (e.g. 1024*1024) images for training. What do I need to do if I train on non-1:1 resolution images? Like dynamic resolution.

[Huage001]

Thanks for the question!

I have 2 ideas for this if you want to train on non-1:1 images.

The first one is a simple one. You can directly remove the resize operation in the codes, to load the original resolution. But the batch size can be only 1 in this case.

The second one is to consider merging specialized image loader such as here. You need to do some alignment on the data format.