higher resolution generate
I try to generate with higher resoltion more than 1536, and the likeness of result quite not good compare to lower resolution,a background still look same but character pose is seem different. Have anything can solve this problem?
Good observation, and it is expected rather than something you did wrong. The LoRA was trained at 1024, and the base Krea 2 model itself was trained around 1 to 2MP, so once you push past roughly 1536 you are outside the range the base knows well, which is when identity loosens and pose drifts. For now the reliable recipe is to generate at 1024 to 1536, then upscale separately if you need more pixels. Native higher-resolution training is exactly what I want to try next, so this should improve. Thanks for the clear report.