Test other LR images

[choieq]

Hi, Thank you for your work.

I wonder, have you test this model on other LR images?
Because when i test my own LR images(LFW), the result was strange.

I convert from HR to LR using your code.

But, unlike CelebA, LFW's resolution is 128x128. So, i remove center crop code.

Result:

Can you advice me? Thank you.

도움주시면 감사하겠습니다! :ㅇ

[DeokyunKim]

Thank you for your interest in our work.
The quality of super-resolved images generated by trained-model relies on train data distribution. i.e., Input low-resolution images has same or similar distribution with train LR-HR pairs(CelebA, LFW).

I think your own dataset has different distribution with our train dataset of coursely. The distribution means noise, blurry, brightness, camera parameters, and etc on your dataset.

In this work, we use downsampling of CelebA and LFW instead of real world LR-HR pairs. Therefore, you should retrain our nerwork with on your own dataset for applying other application or conducting your experiments.

If you have more questions, please let me know. Note that, we can not open overall source code because of company project sequrities.

Thanks!

[choieq]

Thank you for your fast reply!

But above LR examples are in LFW Dataset.
Did you also use LFW to train your network? I know that your network only learned from CelebA.

If your network trained by using CelebA, LFW, shouldn't it work in LFW data samples?

[DeokyunKim]

Did you tried to use "checkpoints/unalign_trained_generator_checkpoint.ckpt" ?

[choieq]

if the test data is celeba -> generator_checkpoint_singleGPU.ckpt
elif lfw -> unalign_trained_generator_checkpoint.ckpt

Is this right? Can you explain other pre-trained weight?
I tried only use generator_checkpoint_singleGPU.ckpt.

It works to use unalign_trained_generator_checkpoint.ckpt in LFW.
Thank you very much your reply.. :)

[DeokyunKim]

compressed_model_011000.pth is weights of Compressed Face Alignment Network.
generator_checkpoint.ckpt is weights of FSR network trained with aligned dataset (CelebA) and Distributed learning.
generator_checkpoint_singleGPU.ckpt is weights of FSR network trained with aligned dataset (CelebA).
unalign_trained_generator_checkpoint.ckpt is weights of FSR network trained with unaligned dataset (LFW).