Closed Davinccci closed 5 months ago
Hi Davinccci,
You are right there is a problem with this denoiser. I am going to delete this version from the drive. Once I investigate what's wrong with the model, I might upload another 64x64 model.
Sorry for the inconvenience.
Charles Laroche
Thank you for your reply, I look forward to your update!
It should be good by now, the new pre-trained model is in the drive.
I did not have the time to do extensive hyper-parameters tuning so you might want to re-tune the parameter controlling the strength of the regularization ($\lambda$) depending on your use case.
Note that this parameter is different for the 64 model than for the 33 model: https://github.com/claroche-r/FastDiffusionEM/blob/7f161175b3b440378581c69df67a9bae0a30fe2a/guided_diffusion/gaussian_diffusion.py#L687-L692
Don't hesitate to tell me if you run into another bug.
Thank you for the update. The model now runs effectively. In fact, during runtime, I found a mismatch between the parameters of the 64 model and the predefined FdnCNN network. Upon comparing parameters, I discovered that the model's network should be the UNetRes model. After replacing it, the model runs smoothly.
Also, when I use your code for the FFHQ 64*64 deblurring experiment, I find that some images are restored very well, while others fail to restore. Is this normal?
Thank you again for your prompt responses and updates.
Great work! However, when I tested using the kernel_denoiser_64.pth model you provided, I found that the blur kernel generated after regularization is incorrect (as shown in the first image). This issue does not occur when using the kernel_denoiser_33.pth model (as shown in the second image). Can you explain why this might be happening? Thank you!