volflow / deflow Goto Github PK

Thanks for sharing.
If I have some noise data want to add in with the AIM-RWSR or NTIRE-RWSR, how to start the training with pre-trained deflow model weights?

Hi,

when running test.py I run into ModuleNotFoundError: No module named 'utils.ImageSplitter'

Could you provide the missing file?

All the best,

Matthias

Hi,
in my data the difference between LQ and HQ data is really big.
so in order that h() will not indicate from which domain (x, y) it comes from I need to do a very strong down-sampling (x8 is not enough).
you have already shared RRDB_PSNR_x8.pth and RRDB_PSNR_x4.pth pre-trained models, can you share RRDB_PSNR_x16.pth as well?
Thanks,
Mani

I'm looking forward to the DeFlow codes.
Thanks.

Why the conv_shift in your "DeFlow-DPED-RWSR-100k.pth" is 0? And I retrian Deflow on my own dataset, but the conv_shift is 0, too. Is it common? Why does this happen？Looking forward to your reply,thanks.

I am training DeFlow on my own generated data.
During the train I sometimes get a raise due to reaching maximum number of Nans of the NLL (100).
Also, on some of my validation images, I get Nans during training, and also on translate.py (and results are of course bad).
Do you have an idea of the root cause of those Nans, and maybe an idea of how to resolve them?
Thanks,
Mani

Thanks for sharing your excellent work. I have a basic question about the NLL loss function.

By using Change of Variable Theory, we can reformulate the NLL p(x) = p(z) * abs(det(inv(Jacobian))), and I did found calculating the second term in your code. But I have a question about whether your did calculate the first term on the right of the equation.

The problem is basic, but I have been confused about it for quite a long time. Do we need to calculate the probability of z (in theory, we need to calculate the inverse of x and find the corresponding z which obeys N(0, I). And we need to calculate the probability of sampling such exact z)?

If we need to calculate the term, could you please tell me which period of code you used?

Looking forward to your reply.

hello author, thank you for your contributions. I have a question about the process of sample in your work. you assume that Zx ~N(0,I), and Zy = Zx + u. So in the process of sample, Can I directly let Zx ~N(0, I) rather than encoding x. I am looking forward to your reply.

Hello
How are you?
Thanks for contributing to this project.
It looks that this method did NOT consider out-of-focus blur or motion blur as low-quality.
Is it possible to use this method for low-quality blurry image generation?
if so, could u guide me how to train a DeFlow model on unpaired dataset consisting high-quality images and low-quality images?
In fact, I looked at your code for training a DeFlow model on my custom dataset.
But the configuration file looks likely to be used for training image super-resolution model rather than simple high-quality-to-low quality generator.

Hello
How are you?
Thanks contributing to this project.
I am going to train a DeFlow model on my custom dataset consisting of UNPAIRED low-quality images and high-quality images.
Could you provide templates of config file and dataset file?

Thanks for your work. But now the pretrained models of RRDBnet can't be got from the website.