Questions about unfrozen new_vae

[mst6750623]

Thanks for your great work！
I noticed there is a new_vae in generator.py and the encoder is unfrozen.

 # copy a new vae
 self.pipe.new_vae = deepcopy(self.pipe.vae)
 self.unfrozen(self.pipe.new_vae.encoder, type(self.pipe.new_vae.encoder.conv_in))

Could you please explain why the new_vae encoder was unfrozen here, as this approach differs from what was described in the paper? Additionally, does this adjustment contribute to improvements in the final results?

[Feynman1999]

Our idea was to fine-tune the VAE for the prior noise stream, so that the distribution of vae_encoder(lq) is closer to that of vae_encoder(gt), which might allow the prior to be utilized more effectively（Especially for the degradation caused by realesrgan noise—during base-model training, the ground truth (GT) may not contain strong noise.）. However, experiments show that this variable has almost no impact on the results.

We did not revert this change in the code, so there is indeed a slight discrepancy compared with what is described in the paper.

For the control noise stream, it should remain frozen, since this helps preserve the possibility of high-fidelity reconstruction.