MrDeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
SPTTecoGAN super resolution neural network
#1
You are not allowed to view links. Register or Login to view.

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

Hello, since it is quite uncommon to have faces with 256 resolution (or more), why not use this algorythm on 128 res faces, and then convert and merge into a video as usual ?

It is far from my knowledge to use it and integrate it into DFL, but I'm currently using JavPlayer, a program to decensor mosaic in japanese porn, which uses TecoGAN successfully. From what I understand it can boost resolution of anything up to 4x, and it would be way faster than training a real 256 or 512 model. From my experience with decensoring japanese porn, using this for deepfakes would take the normal amount of computing time + more or less 20%.

To be clear, contrary to the Topaz AI gigapixel method, we would use this only on faces (like JavPlayer uses it only on the mosaic censored part of a video) and not the entire video. That's why it doesn't take much time. Only after the TecoGAN upscaling is done, the final video would be created.

So if someone can inform the maker of DFL about it, I think it could be awesome. Or maybe there's already something like this in the latest version, I haven't tried DFL since 6 months.
#2
Looks very promissing. Totaly agree with SPT. The results are amazing!
#3
This is very impressive indeed. Hope @iperov see's this, curious to see what he thinks.
~ Fake it till you make it ~
You are not allowed to view links. Register or Login to view.
You are not allowed to view links. Register or Login to view.
#4
DFL already does this mate. Can't remember the exact name, but pretty sure the option is 'Super Resolution' while converting.

Although the model it uses isn't great (DCSCN).
#5
DFL now uses RankSRGAN instead of DCSCN which is sort of like ESRGAN and similar to TecoGAN as they all are based on generative adversarial networks.
Comparison between SRGAN, ESRGAN and currently implemented RankSRGAN: You are not allowed to view links. Register or Login to view.
As you can see RankSRGAN gives best results.
Difference as I understand is that Teco has some temporal coherency capability which means it's better for moving footage, not sure however if it has much of importance for just upscaling a face.
RankSRGAN can give good results provided your model produces sharp enough results, you can always enable box/gaussian sharpening and use that to make super resolution effect more pronounced (I often go as high as 5 or even 8 with box or gaussian if I want a bit "softer" sharpening).
If your predicted faces are blurry super resolution won't make much of difference.
TMB-DF on the main website - You are not allowed to view links. Register or Login to view.
Want to request a paid deepfake or have any questions reagarding the forums or deepfake creation using DeepFaceLab? Write me a message.
Donations (BTC): 1H2vT4saNyULYtm42v2Wm83VuuWHTV4PHW
#6
yes, all super resolution networks require a details.

Currently I am working on new model based on NVLabs\FUNIT
almost zero morphing
highest details
I called it "TrueFace"


[Image: sQUZnTQ.jpg]
consider a donation for next deepfacelab update
#7
A new model... well looks like maybe it was worth to break the SAE model, this looks interesting but it looks like it breaks easily with badly aligned faces (3rd one) and some expressions are kinda weird (looking at eyes, heavily cross eyed in the 1st and 2nd one) but maybe it's just this RDJ dataset.
TMB-DF on the main website - You are not allowed to view links. Register or Login to view.
Want to request a paid deepfake or have any questions reagarding the forums or deepfake creation using DeepFaceLab? Write me a message.
Donations (BTC): 1H2vT4saNyULYtm42v2Wm83VuuWHTV4PHW
#8
break SAE? all works fine.
#9
Now is the era of truefake, not deefake?
#10
(09-15-2019, 04:12 PM)iperov Wrote: You are not allowed to view links. Register or Login to view.yes, all super resolution networks require a details.

Currently I am working on new model based on NVLabs\FUNIT
almost zero morphing
highest details
I called it "TrueFace"

So I only need an handfull of source images and no thousands?

Forum Jump:

Users browsing this thread: 1 Guest(s)