I am wondering whether it is possible to use standard SD model like SD 1.5 SDXL and fintune it using various picture resolutions. Also I am wondering whether I can do that when training a ControlNet model for those underlying SD models.

To my current understanding those standard models only accept square input picture formats like 512×512 and I did not succeeded with inputing for example 768×512, which makes sense, because of size of input layer, right? Is it possible to modify that architecture in some simple way or would that require completely different model architecture to successfully train model based on those underlying ones?

Thanks for any toughts!