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.

H128 vs SAE

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
I've been playing around with SAE and finally was able to produce a model that wasn't a total mess. This was used in this recent video:

https://mrdeepfakes.com/video/2733/...-intern-day2-masturbation-blowjob-and-swallow

Before I tried training SAE - DF, but the tones were off, and certain angles were awful. I then tried SAE - LIAF which worked pretty well.

Here is a frame using H128 vs SAW-LIAF. These are different scenes, but same model, same video.

yVv5qrBh.jpg

Left = SAE-LIAF with overlay - ephc ~ 78,000
right = H128 with histomatch - epoch ~138,000

I'm pretty happy with the SAE model, but it takes maybe 25-30% longer to train.
 

Cerciusx

DF Pleb
How does h128 stand up to the faceswap git original model? I had a few modeld trained with the original and they look better than the lowmem model but I'm guessing since h128 exists its better?
 

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
I think the H128 would be equivalent to the hi-def faceswap model (I forget what it's called). This is only regarding image quality. I have been told post-conversion the models look different and H128 looks more like src? But I cannot confirm this myself without doing tests on the same scene.
 

Cerciusx

DF Pleb
dpfks said:
I think the H128 would be equivalent to the hi-def faceswap model (I forget what it's called). This is only regarding image quality. I have been told post-conversion the models look different and H128 looks more like src? But I cannot confirm this myself without doing tests on the same scene.

I was under the impression that the original model was 128 pixels as well, unless I'm missing something.
 

avalentino93

DF Admirer
SAE is being touted as the far superior model for 2019.   Im just going through my first trainings now.   I can get a good result on H128 now, pretty much every time.  But I wanted to see what this new hubbub is all about.  Anyone else have luck switching to SAE?
 

ZeroCool22

DF Vagrant
LIAEF128 still outperforming SAE in a lot of ocasions, even with all the features SAE have.
With LIAEF the SRC face is a lot more visible, with SAE sometimes it's hard to tell who's the SRC face, even if you leave all at 0 value.
 

iperov

DF Enthusiast
Developer
again
SAE contains H128 and LIAEF.
Disable multiscale decoder and style losses, leave dims default, choose archi DF and HALF face for H128
and archi liae for LIAEF
 

avalentino93

DF Admirer
I find myself messing around with conversion settings for hours.  Do you have any recommendations on best way to convert SAE?  Or any tips?  Its the last part of the puzzle that always stumps me.  The differences in output seem vast between settings.  I realize it depends on x,y,z, but any baseline recommendations to follow?
 

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
avalentino93 said:
I find myself messing around with conversion settings for hours.  Do you have any recommendations on best way to convert SAE?  Or any tips?  Its the last part of the puzzle that always stumps me.  The differences in output seem vast between settings.  I realize it depends on x,y,z, but any baseline recommendations to follow?

For me, OVERLAY works the best. I also color transfer with RCT, and as for other stuff, usuallyt default except for blur. I might use 100-150 in overlay mode.
 

pvogerr

DF Vagrant
ZeroCool22 said:
LIAEF128 still outperforming SAE in a lot of ocasions, even with all the features SAE have.
With LIAEF the SRC face is a lot more visible, with SAE sometimes it's hard to tell who's the SRC face, even if you leave all at 0 value.

I find LIAEF128 works best too specially with side shots
 
I found SAE to be much better than the H128 with the models I have tried out but it is slower but I'd rather have the quality.  I haven't used the LIEF yet, I always use DF, 128 resolution, default dims, and turn on multiscale decoder and learn mask and everything else I keep off.  I never use pixel loss or ca weights or mess with face and background or pretrain a model.  

For conversion, it will depend on the video but I always use overlay, blur changes the most and I start with 30 and go as high as 120. I always check the first few frames to adjust to get rid of the line in the forehead, RCT or LCT both are good since the last LCT upgrade has made it almost even but it will depend on the video.  The other settings I don't mess with.  

I am not an expert but think the SAE is amazing, I am blown away on how real some of the results have been.
 
Top