Training trouble

[tianqyun111]

when training my own in-context-lora,i use some images shown below,but when inference,it usually appear black edges above the picture and at the junction of two pictures. Why? Can anybody give some guidence?

[flankechen]

hi，maybe origin flux 4 grids would product the edging sometimes. you dont want the grid edge when inference?
could you check your kohyass branch in ai-toolkit? you just git clone the main branch of kohyass and runs?
I encounter some error pointing to kohyass like:

AttributeError: 'LoRASpecialNetwork' object has no attribute 'loraplus_unet_lr_ratio'