LoRA training was slow, canceled it and then BSOD. Retried training and it's a fast process again?

LarryJane491 / Lora-Training-in-Comfy

This custom node lets you train LoRA directly in ComfyUI!

364 stars 50 forks source link

LoRA training was slow, canceled it and then BSOD. Retried training and it's a fast process again? #16

Open VGF64 opened 7 months ago

VGF64 commented 7 months ago

Hello,

The LoRA training node works great, however, there were two separate occasions where the training is abysmally slow. I had to cancel on both of them and they blue screened my PC. After all of that, I retried the training and it's a stable, fast process again. Why does this happen?

LarryJane491 commented 7 months ago

Hey there. I have absolutely no idea ^^'. In these two instances, were you using new parameters in the node? Could you try to pinpoint which parameters make it happen?

VGF64 commented 7 months ago

Hey there. I have absolutely no idea ^^'. In these two instances, were you using new parameters in the node? Could you try to pinpoint which parameters make it happen?

There were next to no changes in the parameter except a little change in the training model. What's strange is that one of the checkpoint models I have used with no issue is mainly for realistic people, the SD 1.5 base model. Note that I'm not using the advanced version of the node.