DeokyunKim / Progressive-Face-Super-Resolution

Official Pytorch Implementation of Progressive Face Super-Resolution (BMVC 2019 Accepted)
260 stars 60 forks source link

Unaligned results #12

Closed Steve-Tod closed 4 years ago

Steve-Tod commented 4 years ago

Hi, do your pre-trained models support unaligned inputs? I test it with unaligned inputs and got really bad results. But with aligned inputs, the results look great. image

I found you show unaligned results in your paper, can you please provide the pre-trained model on that set? Thank you!

Steve-Tod commented 4 years ago

The picture above is test result of unaligned CelebA faces with generator_checkpoint_singleGPU.ckpt.

DeokyunKim commented 4 years ago

The trained model we uploaded is trained by aligned face images.

Steve-Tod commented 4 years ago

Thanks, can you please also upload the model trained with unaligned ones?

DeokyunKim commented 4 years ago

I uploaded 'unalign_trained_generator_checkpoint.ckpt'. However, since we preprocessed the unaligned dataset, there is a possibility that the result images do not look good.

Steve-Tod commented 4 years ago

Thank you so much! Another thing I found interesting is that when running your evaluation code, different batch size gives different PSNR result.