I'm trying out a new way of training LoRAs, with a 1400x1400px res + buckets instead of the standard 1024. Dataset composed exclusively of high-resolution images. I can generate up to 1300 x 2000 without any distortions problems.