MrDeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
HaDenGUp-scaling the frames and removing blur using Topaz A.I. Gigapixel
#41
Seconding request for a PM with a link to 4.2.x in order to avoid crashes.
#42
New version out. 4.4.2. But here is all of them.
You are not allowed to view links. Register or Login to view.
#43
Used the app for removing blur without scaling.
You are not allowed to view links. Register or Login to view.

Check the video on the site if you would like to see the end result I also included the mega.nz link for uncompressed one.
#44
I am (still) convinced the best way to go about this is to train the SRGan upscaler on your source footage data, it is just a perfect scenario for it's use!
1. Source footage is usually very high quality
2. This has to be scaled down for training
3. When it comes time to convert we are using SRGan or Gigapixel which is trained on an enormous amount of random data, but we have the high-resolution datasets perfectly matched to our project already!

There's no way in hell I'm ever going to get time to look at training methods for SRGan at the moment.. I mean it is possible, but not likely but I am utterly convinced it would be the next big step for deepfakes.
#45
(08-29-2019, 12:39 PM)creepfake Wrote: You are not allowed to view links. Register or Login to view.Hey this is a good idea; it looks like there's a lot of upscaling code out there and I had a look around for something free that would work similarly. The most promising thing I found is using TensorLayer:

You are not allowed to view links. Register or Login to view.

And this is the specific upscaling script example that's in the examples:

You are not allowed to view links. Register or Login to view.
You are not allowed to view links. Register or Login to view.

Haven't set it up but I'll try to see how well it works.

For some examples, here's the SRGAN project before it was integrated into the above links:
You are not allowed to view links. Register or Login to view.

As per this post; SRGAN on tensorlayer seemed like the best solution to me. I didn't get the actual upscaling working but you can install the AI Machine by installing Ubuntu (18.04 LTS) and issuing roughly the following series of commands:

Code:
sudo apt update
sudo apt upgrade
sudo apt install python3-pip

#cuda
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-ubuntu1804.pin
sudo mv cuda-ubuntu1804.pin /etc/apt/preferences.d/cuda-repository-pin-600
sudo apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/7fa2af80.pub
sudo add-apt-repository "deb http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/ /"
sudo apt-get update
sudo apt-get -y install cuda

#specific pre-reqs + tensorflow + tensorlayer
pip3 install pillow
pip3 install easydict
pip3 install scipy==1.1.0    
pip3 install tensorflow-gpu==2.0.0-b1
pip3 install git+https://github.com/tensorlayer/tensorlayer.git
And then run the script at the github\SRGAN link, which I did not figure out yet.

Got it to run in Windows, but it's more likely to get Out of Memory errors since Windows uses so much RAM
#46
thanks, I was just looking for something like that, and
P.S. fake looks gorgeous
#47
For those of you doing this. You should do it on 30 fps, and then interpolate the rest of the frames back to 60 after it is done in sony vegas.

Here is a tutorial, ive tested this and the results are good.  

You are not allowed to view links. Register or Login to view.

Edit: Wait...you don't even need this plugin to re sample to 60fps in Sony Vegas. You can even go up to 120+ without it. opps
#48
For anyone interested you can run this script I talked about before native in Windows - you need to install a Python 3.7 environment and some special prerequisites or versions of certain Python programs. Most errors are covered in the issues section of the github:

You are not allowed to view links. Register or Login to view.

I got the program to run in Windows but since Windows monopolizes so much memory, it is prone to memory crashes and you might need to change batch size or something. I got to the point of running the script but would need to rewrite certain parts of the script, which I wasn't down for. You can easily change the script to accept .jpg files which is helpful, but it seems hard to produce results without more effort and I suspect some better SRGAN program will come along.
#49
I've used Topaz on lower res videos I have that I need to clean up. I can't remember the resolution, but I think the original was 1024 x 768.

Original video before any AI upgrade: You are not allowed to view links. Register or Login to view.

Same video, only enhanced (no upscale): You are not allowed to view links. Register or Login to view.


Enhancing lower res videos made a big difference. Lower res also much faster to AI convert (but still took 4-5 hours total). I still have a dream to AI enhance and convert that old True Anal Stories 9 video with Meridian and Gabriella Kerez! The recent Katy Perry video with Gianna Michaels is also a lower res, low quality video which was AI enhanced. Topaz AI for video just public released today (was in beta previously) .

Forum Jump:

Users browsing this thread: 2 Guest(s)