The training details about Unet parameters frozen

[hda-xian]

Why are the UNet parameters frozen during training for SD1v5, but not for SDXL? the haggingface training sdxl script sets " Unet.train() "

***** huggingface train sdxl *********
vae.requires_grad_(False)
text_encoder_one.requires_grad_(False)
text_encoder_two.requires_grad_(False)
t2iadapter.train()
unet.train()

*******Tencent ARC train sd1v5 **************
model.cuda()
model.eval() # model is contain all models vae ,cliptext
return model

*******Tencent ARC train sdxl **************
vae.requires_grad_(False)
text_encoder_one.requires_grad_(False)
text_encoder_two.requires_grad_(False) -> the Unet does not set no grad means Unet need grad

[bonlime]

because they have params_to_optimize = adapter.parameters(), and only optimise the adapter's params. But disabling unet grads reduces the memory consumption, they just forgot it

[hda-xian]

这是来自QQ邮箱的假期自动回复邮件。你好，我最近正在休假中，无法亲自回复你的邮件。我将在假期结束后，尽快给你回复。