Trained three times, at 512x512, 512x768, and 768x512 (identical data, different crops), and then merged with all three models at 100% strength, in Kohya (which is why LoRA strength has to be at 0.1).
Trained three times, at 512x512, 512x768, and 768x512 (identical data, different crops), and then merged with all three models at 100% strength, in Kohya (which is why LoRA strength has to be at 0.1).