The original IP2P is trained on 256x256 images.
Is it possible to train IP2P on a non-square aspect ratio like 768x512 (3:2)?
There is some hard-coded squareness in edit_dataset.py and the config but does it is it a restriction of the architecture or does it only have to do with the preprocessing code and so on?
The original IP2P is trained on 256x256 images. Is it possible to train IP2P on a non-square aspect ratio like 768x512 (3:2)?
There is some hard-coded squareness in
edit_dataset.pyand the config but does it is it a restriction of the architecture or does it only have to do with the preprocessing code and so on?Thanks!