search

TencentARC / GFPGAN

GFPGAN aims at developing Practical Algorithms for Real-world Face Restoration.
Other
35.86k stars 5.94k forks source link

results are strange while testing on whole img #29

Closed tachikoma777 closed 3 years ago

tachikoma777 commented 3 years ago

Hi, I'm using pretrain model for testing, while testing on align cropped faces, the results are promising. However, while testing as whole img the results are wrong, seems the detection is not working well because in results cmp folder there is no faces or sereval faces in one img.

I didn't change the code, is there anything wrong the whole img testing? Thanks in advance.

xinntao commented 3 years ago
  1. If no face is detected, it will restore it.
  2. As the current model also enhances colors, when putting faces back to the whole image, the results are not good. We will release a model w/o changing colors.
tachikoma777 commented 3 years ago
  1. If no face is detected, it will restore it.
  2. As the current model also enhances colors, when putting faces back to the whole image, the results are not good. We will release a model w/o changing colors.

But I use default test img in original project, it will be strange if no face is detected. results are as follows: image

tachikoma777 commented 3 years ago

By the way, I have i small question about the Channel-Split SFT, which is how the F_prior is splited in chanels? How many channel F_prior have? Thanks!

xinntao commented 3 years ago
  1. If no face is detected, it will restore it.
  2. As the current model also enhances colors, when putting faces back to the whole image, the results are not good. We will release a model w/o changing colors.

But I use default test img in original project, it will be strange if no face is detected. results are as follows: image

Did you use the colab demo? It seems that there must be something wrong with face detection.~

xinntao commented 3 years ago

By the way, I have i small question about the Channel-Split SFT, which is how the F_prior is splited in chanels? How many channel F_prior have? Thanks!

Half the original channels~

tachikoma777 commented 3 years ago
  1. If no face is detected, it will restore it.
  2. As the current model also enhances colors, when putting faces back to the whole image, the results are not good. We will release a model w/o changing colors.

But I use default test img in original project, it will be strange if no face is detected. results are as follows: image

Did you use the colab demo? It seems that there must be something wrong with face detection.~

No, I tested on local side. After I recheck the crop region and found that for every input the detection region is always the left up 512*512 block. I tested on both moblienet and resnet detection model. Is there something wrong with facexlib? I tested both v0.1.3.1 and v0.1

xinntao commented 3 years ago

I have updated the model without colorization. Colab demo: https://colab.research.google.com/drive/1sVsoBd9AjckIXThgtZhGrHRfFI6UUYOo

For your problem, I have not encountered this issue on my side.