about scale factor

[shiyuleixia]

Super-resolution settings

basesr:
sf: 1
chopping: # for latent diffusion
pch_size: 512
weight_type: Gaussian

I change the scale factor to 1,but it dosen't work:

[zsyOAOA]

Currently, our model only supports x4 SR task.

[flynnamy]

How to modify this? @zsyOAOA

[sunzhe09]

+1

[zsyOAOA]

@sunzhe09 @flynnamy I have stated that our model only supports the x4 super-resolution task recently.

[flynnamy]

How to modify this? I do not understand why the scale factor has been fixed? @zsyOAOA

[zsyOAOA]

How to modify this? I do not understand why the scale factor has been fixed? @zsyOAOA

Our noise predictor is trained to predict the noise map in the latent space of VQGAN, which is downsampled 8x times spatially. Thus we have to adjust the number of downsampling operators according to the scale factor. E.g., for x4 SR task, the noise predictor contains two downsampling layers as shown in [config[(https://github.com/zsyOAOA/InvSR/blob/master/configs/sample-sd-turbo.yaml#L43) to match the resolution of latent space. Therefore, we can't directly modify the scale factor! @flynnamy

[ManuBN786]

Thant means even if I put sf:1 during training, I still cant generate 1X SR at prediction?