Gigapixel Face Refinement vs Remini

Hi, i am wondering if you will offer other AI methodes in the future. Rimini Photo app is more “inventing” new face information, based on their models rather then pure enhancing.

Would be nice to see you guys having the first Desktop ready Solution. Rimini uses quite some Serverpower to make their service aviable on mobile phones.

1 Like

You are far from being the first to mention Remini here - there are several topics on it already.

1 Like

Hi PaulM, i used the search function and nothing came up searching for “Rimini”.
Ok, now i see all ot the threats. Thank you.

1 Like

The problem is that a smartphone doesn’t have as much power as an entry-level GPU and has to send the images to the server.

And then the images belong to someone else.

1 Like

There are funny examples
Actually the technology is fundamentally different in the way that Remini has used machine learning, its doing something that is new

For professional use Remini has issues.

  1. it concentrates only on the face areas and uses different technology to retrieve definition in face, than the rest of the image. leading to weird transitions between the areas. The face areas can be frankly in a different league. at one point it was capable of producing skin texture that was extraordinarily realistic, Hair - not so good.
  2. They keep shifting the results so its not reliable.
    It’s worth the subscription though, and better value than GigaPixel now sadly
1 Like

I found that there are many open-source face upscale AI code in Github. And their result are even better than Remini and MyHeritage. :heart_eyes: Just curious, why Topaz lab don’t “borrow” some open source Face Reconstruction AI algorithm into Gigapixel. :thinking:

I have tested your image using Face Restoration AI created by Tencent ARC.
And here is the result. :yum:
Clearly the result is much better than GAI’s Face refinement. :sweat_smile:

2 Likes

You clearly haven’t noticed that was December 2020. What you are showing there is dreadful color and the eye reconstruction looks false as do the eyebrows, but the current version of GigaPixel does a good job using the very compressed model:

Here is another example using Tencent ARC’s GFPGAN :grinning:


Left: Original Image
Right: After Face restoration

Compare with GAI :yum:


Left: GAI compress model Face ON
Right: Face Restoration by Tencent ARC’s GFPGAN

Original Image: 110 px x 110 px
original

1 Like

The main reason why topaz’s solutions aren’t as good is that these other programs/services rely on the cloud and a much, much larger database, while topaz’s progs are local as there’s a much more finite source of face examples to choose from due to limited HDD space, whereas the cloud is capable of producing much much more!

It’s like noting the different varieties of cereal in a small convenience store vs Cosco…

Not really, because you still have to rent it over an extended period of time, whereas with Topaz products you pay a perpetual license!

The Tencent ARC’s GFPGAN, I last tested doesn’t require internet connection.
The Pretrained AI Model “GFPGANCleanv1-NoCE-C2.pth” I used in the test, is only 332 MB in size.

1 Like

I stand corrected… Whelp, at least VEAI has a goal to shoot for!

this Tencent ARC’s GFPGAN is very interesting, how can I use it? I dont see a way to have an installer of sorts, does it have to be deployed on a server in order to run it?

You have to install Python in your computer and clone all the files to your computer.

Then install all the required stuffs listed in the requirements.txt.

Then download their latest pre-trained AI model from Google Drive.
https://drive.google.com/drive/folders/17rLiFzcUMoQuhLnptDsKolegHWwJOnHu?usp=sharing

After all the stuff installed correctly, you can use Command Prompt to run the script. Such as,

python inference_gfpgan.py --upscale 2 --test_path inputs/whole_imgs --save_root results

And the result I tested is pretty amazing. :grinning:

1 Like

thank you, crunching through installations of many kinds, will get back once it doesnt work, cheers.

Collecting basicsr
  Using cached basicsr-1.3.1-py3-none-any.whl (144 kB)
ERROR: Cannot install basicsr==1.3.1, basicsr==1.3.2, basicsr==1.3.3, basicsr==1.3.3.1 and basicsr==1.3.3.3 because these package versions have conflicting dependencies.

The conflict is caused by:
    basicsr 1.3.3.3 depends on torch>=1.7
    basicsr 1.3.3.1 depends on torch>=1.7
    basicsr 1.3.3 depends on torch>=1.7
    basicsr 1.3.2 depends on torch>=1.3
    basicsr 1.3.1 depends on torch>=1.3

To fix this you could try to:
1. loosen the range of package versions you've specified
2. remove package versions to allow pip attempt to solve the dependency conflict

ERROR: ResolutionImpossible: for help visit https://pip.pypa.io/en/latest/user_guide/#fixing-conflicting-dependencies
WARNING: You are using pip version 21.2.3; however, version 21.3.1 is available.
You should consider upgrading via the 'C:\Users\Coudy\AppData\Local\Programs\Python\Python310\python.exe -m pip install --upgrade pip' command.

it seems I dont know how to install “torch” I though I have successfully installed pytorch through anaconda and also cloned pytorch through github so I dont know what “basicsr” wants.

You may also follow the install instruction from https://pytorch.org/

For example,
CUDA 11.3:

pip3 install torch==1.10.0+cu113 torchvision==0.11.1+cu113 torchaudio===0.10.0+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

CUDA 10.2:

pip3 install torch==1.10.0+cu102 torchvision==0.11.1+cu102 torchaudio===0.10.0+cu102 -f https://download.pytorch.org/whl/cu102/torch_stable.html

CPU:

pip3 install torch torchvision torchaudio

You can also find pervious version from

You may need to find one that fit your system.

all this went smooth and installed correctly, CUDA and CUDnn. Problem is somewhere with path variables I think.
“import torch” gives an error

Downloading basicsr-1.3.3.7.tar.gz (117 kB)
     |████████████████████████████████| 117 kB 3.3 MB/s
  Preparing metadata (setup.py) ... error
  ERROR: Command errored out with exit status 1:
   command: 'C:\Users\Coudy\AppData\Local\Programs\Python\Python310\python.exe' -c 'import io, os, sys, setuptools, tokenize; sys.argv[0] = '"'"'C:\\Users\\Coudy\\AppData\\Local\\Temp\\pip-install-44x8l6p2\\basicsr_f7549a236bda4074a607b526db98e5da\\setup.py'"'"'; __file__='"'"'C:\\Users\\Coudy\\AppData\\Local\\Temp\\pip-install-44x8l6p2\\basicsr_f7549a236bda4074a607b526db98e5da\\setup.py'"'"';f = getattr(tokenize, '"'"'open'"'"', open)(__file__) if os.path.exists(__file__) else io.StringIO('"'"'from setuptools import setup; setup()'"'"');code = f.read().replace('"'"'\r\n'"'"', '"'"'\n'"'"');f.close();exec(compile(code, __file__, '"'"'exec'"'"'))' egg_info --egg-base 'C:\Users\Coudy\AppData\Local\Temp\pip-pip-egg-info-b3gttlwu'
       cwd: C:\Users\Coudy\AppData\Local\Temp\pip-install-44x8l6p2\basicsr_f7549a236bda4074a607b526db98e5da\
  Complete output (5 lines):
  Traceback (most recent call last):
    File "<string>", line 1, in <module>
    File "C:\Users\Coudy\AppData\Local\Temp\pip-install-44x8l6p2\basicsr_f7549a236bda4074a607b526db98e5da\setup.py", line 8, in <module>
      import torch
  ModuleNotFoundError: No module named 'torch'

For all this kind of stuff you better use Miniconda3 and create a specific environment for it to avoid any conflicts with some other python applications

conda create --name gfpgan
conda activate gfpgan

Then you install all the dependencies

conda install python==3.7
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

Great, thanks for both of you on this matter, somehow played around with both and copied things from left to right but now it processes the photos I put in so thats great. I can your Topaz Labs products to prepare the background and combined it all in photoshop. Perfect.
I probably wouldnt be able to replicate how I managed to install the whole thing though. Lots of trial and error.

1 Like

Hello, do you mind me asking about those pretrained models available for download for GFPGAN? I have checked that they are not as high res when upscale 4x. Do you happen to know what is that pretrained model based on? I found some links there to a 70000 Flickr faces database and wonder if I could train a model myself or if I could use my own photo library to create a model specifically for my family blind face restoration?