use sd_dreambooth_extension to training lora, model file is too big

[segatecm]

I try to open the Lora option, and the training result file amost 8G.
But the kohya lora tool just 120M, Please tell me why?

[sailxjx]

Just save the lora, not the checkpoint, and reduce the unet and text encoder rank, 32/64 is enough.

[segatecm]

thanks, but how to it in sd_dreambooth_extension ?

[sailxjx]

In the save tab,
Uncheck Generate a .ckpt file when saving during training.
Check Generate lora weights when saving during training.

[segatecm]

Thanks!