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.

Upscaling resolution of deepfakes.

wbdragon

DF Pleb
Haven't seen anyone mention this here, but there are some repositories on github that use deep learning to upscale resolution of images. 

You can try 5 images for free at https://letsenhance.io/

I have used it to make some 4k wallpapers. lets enhance works well with photos that look 720-1080 ish. But I have seen more impressive examples with the stuff they use in github, because they have some good trained models. You can see these with some googling and youtube. 

https://github.com/topics/srgan

Looks like the up scaling is quick. In lets-enhance, it takes a few seconds for me, with a large image. So I think it could be used with a workflow. 

This would be great if you want to make VR deepfakes, or simply upscale your video.
 

Pocketspeed

DF Admirer
Verified Video Creator
wbdragon said:
Haven't seen anyone mention this here, but there are some repositories on github that use deep learning to upscale resolution of images. 

You can try 5 images for free at https://letsenhance.io/

I have used it to make some 4k wallpapers. lets enhance works well with photos that look 720-1080 ish. But I have seen more impressive examples with the stuff they use in github, because they have some good trained models. You can see these with some googling and youtube. 

https://github.com/topics/srgan

Looks like the up scaling is quick. In lets-enhance, it takes a few seconds for me, with a large image. So I think it could be used with a workflow. 

This would be great if you want to make VR deepfakes, or simply upscale your video.

I've been looking into this lately. I haven't had much free time, so I haven't attempted it, or even cloned the git yet. But I've been looking at the ESRGAN git myself.

At first glance, it potentially looks like a method of increasing resolutions for videos that are not HD. Personally, I would probably use it on videos that are less than 480p. No need to upscale 480p, in my opinion, because most current training models don't use faces greater than 256 pixels anyway.

I have two questions/concerns about it. The first is how well it would work on faces? I've read conflicting reports. The second is how an artificially upscaled face would affect the training results in Faceswap or DFL?
 

Cerciusx

DF Pleb
Tried the site, did not upscale my pictures nor did I notice any clear difference. The photos are the same resolutions so I see no point.
 

wbdragon

DF Pleb
Cerciusx said:
Tried the site, did not upscale my pictures nor did I notice any clear difference. The  photos are the same resolutions so I see no point.

It does work. If you try to make 720p/1080p ish photos to 4k, it gives a great improvement. Lower res photos, it looks like it just went through a filter.
 

Schmitty

DF Vagrant
I use Topaz AI Gigapixel. It's a relatively big install because it's basically a local envelope for all the usual machine learning libraries currently used in AI-based image enlargement. It's amazing in many ways, but you need to be careful about some of the smoothing settings ruining the eyes. However I have successfully used it to upscale videos.
 

wbdragon

DF Pleb
Schmitty said:
I use Topaz AI Gigapixel. It's a relatively big install because it's basically a local envelope for all the usual machine learning libraries currently used in AI-based image enlargement. It's amazing in many ways, but you need to be careful about some of the smoothing settings ruining the eyes. However I have successfully used it to upscale videos.

I just tried the Topaz program. And it was not good. https://letsenhance.io/ is WAY better, but you have to use there service and it would take forever and cost money. 

Topaz was really similar to what Photoshop already does. It's not going to be as good as with something with training data. 

I'm on a 4k display I can really tell the difference. 


Cerciusx said:
Tried the site, did not upscale my pictures nor did I notice any clear difference. The  photos are the same resolutions so I see no point.

Also lets say you have an image at 1080 scale, but it really looks 720. Take that Image in photoshop, and resize it to 720. Then upscale and you should see results.


God I want to figure out how to do this...CHECK THIS OUT!!!

 

Pocketspeed

DF Admirer
Verified Video Creator
@wbdragon You might want to search on Github for SRGAN and/or ESRGAN. SRGAN was used for the Youtube video that you linked. ESRGAN is supposedly an enhanced version of SRGAN. 

I've read conflicting reports on how well either one does with faces.

The documention on both is not very friendly for new users, however. I haven't tried either one yet, but it's on the to-do list. It'll probably be a few months before I get to it.
 

Cerciusx

DF Pleb
if you want to upscale artwork anime/drawings/cartoons rather than photographs try out waifu2x. I have used it recently and have turned shitty low res art into really detailed wallpapers for my p52. good shit. theres an online demo various places where you can upload it and choose how to upscale and denise.
 

HarryPalmer

DF Pleb
Verified Video Creator
wbdragon said:
Schmitty said:
I use Topaz AI Gigapixel. It's a relatively big install because it's basically a local envelope for all the usual machine learning libraries currently used in AI-based image enlargement. It's amazing in many ways, but you need to be careful about some of the smoothing settings ruining the eyes. However I have successfully used it to upscale videos.

I just tried the Topaz program. And it was not good. https://letsenhance.io/ is WAY better, but you have to use there service and it would take forever and cost money. 

Topaz was really similar to what Photoshop already does. It's not going to be as good as with something with training data. 

I'm on a 4k display I can really tell the difference. 


I'm also on a 4K display, and I'm really surprised to hear what you say about AI Gigapixel. I also tried LetsEnhance and Waifu, and AIG is way more configurable. I suggest users take advantage of the 30-day free trial and decide for themselves. Incidentally, there has just been a major upgrade to the program.

 

Kenturin

DF Vagrant
jIuWjdRh.png


To clarify...the first picture started off as a 128 x 128 pic. Second pic ran through an SRGAN upscaled 4x to 512 x 512. I ran both the original and the SRGAN through DFL extractor to produce 256 x 256 aligned extractions. These are the results. YMMV
 

Pocketspeed

DF Admirer
Verified Video Creator
@Kenturin So I'd be interested if there is a step by step guide or tutorial available for SRGAN. Any chance you could point me in the right direction?
 

Kenturin

DF Vagrant
Pocketspeed said:
@Kenturin So I'd be interested if there is a step by step guide or tutorial available for SRGAN. Any chance you could point me in the right direction?

I produced that upscale using this repository.  The walk-through is fairly clear and concise and you'll be looking specifically for the "EDSR + SRGAN" section.  It does take some setting up of python if you don't have it installed already.  I'm not well versed in python so that took some stumbling around the internet to get set up.
 

Nickrus

DF Admirer
It might be a good approach to take a smart sharpening filter for Photoshop, there are lot of them, starting from a standard high pass and going on and on. Then create a batch and let it process the output images. It might be also possible to apply a mask (let it work only in the replaced face area) which should conserve CPU/GPU time.
 
Top