Is this possible, to improve? about gfpgan HOT 10 OPEN

@zelenooki87 Thanks for your suggestion.
We will improve it, but it will take some time~

from gfpgan.

@zelenooki87

It would be great if you could upload those strange results :-)

I have just updated GFPGAN, which support using Real-ESRGAN to enhance the background (non-face) regions. You may have a try. The colab demo has also been updated: https://colab.research.google.com/drive/1sVsoBd9AjckIXThgtZhGrHRfFI6UUYOo

from gfpgan.

Thanks 👍

Your suggestion is good.
I updated a version of facexlib, which removes too small faces or side faces, for failure cases as you provided. More details: xinntao/facexlib@24493fc

But it may be not perfect, as it now hard-codes the eye-to-eye distance as a threshold.
I will update it later， and will also consider your suggestion that modify code if face area is too small or blur that then model not generate output.

Thanks

from gfpgan.

Thank you for improving your code. Respect!
On some of photos if persons are too far and not clear, there are very strange results....could you make minor changes to code to choose which faces are included in final results? like reproduce same image again without strange faces manually....i will upload some strange results if there is need to....Also, do you suggest to use upscale software(real esrgan, or commercial solutions like gigapixel ai) prior or after using gfpgan?
on >90% results are great! thank you

from gfpgan.

Thats great. results are now better. however, due cuda out of memory error, for 10-25MPx resolutions, I am using Gigapixel Ai to upscale photos on PC ... Modified code to not upscale with esrgan which is default now, because of cuda error.. With RTX 3060 12GB I am able to inference GFPGAN at 36-40 MPx resolution on Windows.. results are much, much better when code is used properly upscaled and denoised/sharpened photos....

On few number of photos I noticed third eye or even onother mouth added.... mostly on baby faces
like this....
https://i.postimg.cc/d30nLSjc/0009-gigapixel-standard-width-4954px-Sharpen-AI-focus-00.png

I am sure that with time model will be updated. Keep this project live man. All the best!

from gfpgan.

@zelenooki87

1. For the cuda out-of-memory issue, you can set --bg_tile option
2. Could you please provide the whole image for https://i.postimg.cc/d30nLSjc/0009-gigapixel-standard-width-4954px-Sharpen-AI-focus-00.png
   It seems that there is something wrong in face crop.

from gfpgan.

Another example cropped upscale with real esrgan 4x with face enhance option

And crop from original photo

I must say with every new release of code and updated pypi packages from requirements file(real esrgan gfpgan, basicsr and facelib) results are better. results are even now far better than when upscale with latest Gigapixel AI prior. they messed things whith every new version their models seems to be less quality. But back on topic, I suggest you if you can modify code if face area is too small or blur that then model not generate output(for example 64px and less) like in this example. Cause results are not excepted... I will upload more examples if you need

Big support to you for great work and release code for free to public, good man!

from gfpgan.

example from previous post was with freshly anaconda environment with installed strictly most recent version from https://pypi.org/ argumented with == during adding to env.
packages are:
--conda list
https://pastebin.com/dCNZQNYB

Will constantly check your repositories for updates.... and update them all in new environment. My galery contains around 11k of family photos, from low resolutions like old cell phone captured 640x480pixel to 24Mpx DSLR. And lot of scanned photos. Will inform you if output images has unexpected results.

from gfpgan.

@zelenooki87 Thanks:-)

So many family photos. You have a warm and loving family 😃

from gfpgan.

You have updated code and models. I am getting much more natural results with 1.3 model. Will test further and report on eventual problems. This project is really amazing.

from gfpgan.