file.png (370 KB, 832x1216) >>8155486 The moment I switched to sdxl_train.py, the proper colors emerged with my vpred+ztsnr conversion tests. While full-finetune "lora" fried the shit out of OUT5 layer. It cannot be a coincidence, anon. >Because they're good enough No they are NOT. 8gb braindead poorfags are really hurting this field. Your loras gonna be infinitely better if you full finetune nf4-quanted model then extract the lora which would maybe would require the same amount of vram.
Anonymous 08/19/24(Mon)03:27:49 No.8155497
file.png (370 KB, 832x1216)
>>8155486
The moment I switched to sdxl_train.py, the proper colors emerged with my vpred+ztsnr conversion tests. While full-finetune "lora" fried the shit out of OUT5 layer. It cannot be a coincidence, anon.
>Because they're good enough
No they are NOT. 8gb braindead poorfags are really hurting this field. Your loras gonna be infinitely better if you full finetune nf4-quanted model then extract the lora which would maybe would require the same amount of vram.