MrDeepFakes Forums

Some content may not be available to Guests. Consider registering an account to enjoy unrestricted access to guides, support and tools

  • We are looking for community members who are intested in helping out. See our HELP WANTED post.

Using AI to increase our AI? (lol)

Kamrooz

DF Vagrant
Considering the software we're using is AI based. Was wondering if anyone has tried using AI Gigapixel, ESRGAN, STRGAN, etc....to increase the quality of our training images.

For those that don't know. Ai Gigapixel is a AI upscaling application, just like ESRGAN but MUCH easier to use as it has a gui. Here are some quick samples to compare. Took the alligned faces and just ran them through Gigapixel.

It doesn't do much for higher resolution images (as you can see from samples 3 and 6), but for low res images it does wonders. Recreates detail but still keeping the image as real to the original as possible. Although if you plan to go through this process for all your images, I would recommend double checking as rarely from time to time there are anomalies, although not often.

I run the aligned images through gigapixel, which resizes it to 2x, 4x, 6x, custom ranges...I use 2x. Than you can use whichever batch resizing method you want for the upscaled files to 256x256 to match the original dimensions needed for DFL. This software works wonders for src an dst quality enhancements for training purposes. Just wanted to toss this info out there just in case anyone hasn't thought of using this method.

AI Gigapixel though is not a free software, 30 day free trial though, $100 after for a lifetime license. Where as esrgan and alternative upscaling software out there are free to use, but much more complex to function. I prefer the ease of use of AI Gigapixel.



L69mBijh.png
OHJg7NIh.png

JIqahXwh.png
AcnDcqyh.png

fIjXGm8h.png
XZGPOeVh.png

lgmvMyBh.png
mf6hMWGh.png

D4Jzn06h.png
Bsa8LgXh.png

D7VVdXrh.png
19Hw6EWh.png

XpDshZKh.png
qfIiz1Vh.png
 

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
Check out this DCSCN network implementation from DeepFaceLab.

3CRZFDy.gif
 

HarryPalmer

DF Pleb
Verified Video Creator
I have an like AI GIgapixel, but I must admit I had not thought to use it in this way. Interesting. The only thing I would say is to check that it does not mess the eyes up on upscale.
 

Kamrooz

DF Vagrant
Yea. That's one thing I've noticed. I use it sparingly. mostly to increase the quality of my source files. If the DST is too low of a quality file. I'll also run it on each frame and compare to see if it would be better to train against. But it's hit or miss. But for low quality src files, that aren't beyond help, I've found it SUPER useful.
 
Top