Is the backbone Denoising Unet need training？

[TZYSJTU]

Is the backbone Denoising Unet need to be finetuned? Or it is frozen as the SD 1.5 model?
In the paper, it says, "During this stage (second), we only train the temporal layer while fixing the weights of
the rest of the network.‘’
But I see you said "Freezing unet is not a good choice" ?
So I got confused, please answer this question, thanks！

[guoqincode]

Stage1 requires training and Stage2 is frozen

[XingML]

Stage1 requires training and Stage2 is frozen

But I find that the unet is frozen in the stage 1 in your code. Is that normal?

Open-AnimateAnyone/train_hack.py

Line 259 in 02d4500

unet.requires_grad_(False)