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.

SAE vs. H128

halo2k

DF Pleb
Verified Video Creator
Hi,

can someone explain more detailed what the differences are? On Iperovs GitHub he writes 

"SAE - actually contains all other models, but better due to multiscale decoder + pixel loss. Just set style powers to 0.0 to work as default (H128/DF/LIAE) model."

So if my Hardware is powerful enough, should i Always use SAE?

What does the multiscale decoder do?
What is meant by pixel loss?
What are dimensions? ae_dims ed_dims?
"figure out the max dimensions that you can use given your GPU's Memory" how?
What is "NN"?
What is "style power"?

Taking this one as an example:



I don't see any difference, what makes SAE better?

Sorry for the tons of Questions, jus trying to learn, improve and understand this Technology.

Thanks to anyone for answers and explanations.
 

zxz003

DF Vagrant
Bump. wanted to ask the same question. Where can I find the specific detail on introduce each training model?
 

Cerciusx

DF Pleb
really what it comes down to is your face set and your pornstar match. peoplke get all in a huff over these new encoders and model formats but really the best fakes have been made with the default settings of the git faceswap repo and fakeapp. just look at the earliest videos. They are great. The reality is that the face detector library is only going to be so perfect. new model typoes arent going to pick up new angles or give you extra frames. spend time on your faceset and find a good donor body. None of the big deepfakers are using these fancy new model types. I bet you anything.
 

halo2k

DF Pleb
Verified Video Creator
Well that's nothing i don't already know + the answer doesn't help at all.

To be able to use something efficiently and improve, you first have to understand it. Also i bet Iperov wouldn't include it, if there wouldn't be any benefits over e.g. H128 at all.

There are a few good fakes yes, but 95%+ is pure crap. Not matching bodies, too less trained + wrong aligned faces.
Heck in most fakes you can't even recognize the celebrity, where you would wonder why it even has been uploaded.

So if anyone has the knowledge, please share it. I guess it is not only me, being interested in learning and improving.

Thanks in advance.
 

Cerciusx

DF Pleb
So you think that a new model is going to make the pornstars bpdy look more like the celebs and reduce the need for training and magically align faces.... riiiight. If you know this already then shush and improve your facesets. you are on this site a decent amount which means that 95 percent of the current fakes are indeed NOT crap otherwise you wouldnt bother. Nobody here uses fancy new models. the main benefit of differing models is to cater to the spec of the card and workflow.
 

halo2k

DF Pleb
Verified Video Creator
Did you even read my text? Did i wrote anywhere that i think a new model would change a pornstar body? NO! So WTF, don't make anything up.

All i asked, and still want to know is, what these settings do. I never wrote anywhere that i expect magic to happen, nor perfect fakes nor anything.

And indeed most fakes are crap. I could instantly post you several screenshots taken from random videos and hands down, you couldn't even tell me, which celebrity it should be.

So, returning to the inital request, if someone is capable of explaining, please do so.
 

Cerciusx

DF Pleb
Youy said they were crap because of poor training, badly aligned pics and bad body choice, so why not start there before trying a new model? you might get good results like most of us do.
 

frosty3907

DF Admirer
Verified Video Creator
In my experience so far the SAE model is retaining far more of the source face but at the expense of getting a usable conversion.
I have no idea if it's accurate or not but the way I think of it is that SAE is taking your src face and morphing it to match dst, as opposed to morphing dst to try and look like src.
 

halo2k

DF Pleb
Verified Video Creator
frosty3907 said:
In my experience so far the SAE model is retaining far more of the source face but at the expense of getting a usable conversion.
I have no idea if it's accurate or not but the way I think of it is that SAE is taking your src face and morphing it to match dst, as opposed to morphing dst to try and look like src.

If you read the third line of my initial post, this is not how you think it works. From this line (initial post), it reads like SAE is the better H128 as it contains this multidecoder + pixel loss on top.


Cerciusx said:
Youy said they were crap because of poor training, badly aligned pics and bad body choice, so why not start there before trying a new model? you might get good results like most of us do.

Here you go, please name the celebrities:

 

frosty3907

DF Admirer
Verified Video Creator
Yeah man I didn't say I knew how it worked but rather the results I was getting.

It's damn near impossible to get a description of how the models work beyond: "hidden layers" that makes sense to anyone not in the ML field.
 

halo2k

DF Pleb
Verified Video Creator
frosty3907 said:
Yeah man I didn't say I knew how it worked but rather the results I was getting.

It's damn near impossible to get a description of how the models work beyond: "hidden layers" that makes sense to anyone not in the ML field.

Wasn't meant to be rude, don't get me wrong.

I will wait for Iperov to return.
 

Cerciusx

DF Pleb
You missed my point. The reason for crappy deepfakes are various, but it is not the default models fault. train more, better image source etc ... these can all be improved for a better result. The default model was designed for this. Its worth a shot improving your workflow before you go ham on a new model, or you wont know what affected the result, better photos/training or a different model.
 

halo2k

DF Pleb
Verified Video Creator
halo2k said:
Hi,

can someone explain more detailed what the differences are? On Iperovs GitHub he writes 

"SAE - actually contains all other models, but better due to multiscale decoder + pixel loss. Just set style powers to 0.0 to work as default (H128/DF/LIAE) model."

So if my Hardware is powerful enough, should i Always use SAE?

What does the multiscale decoder do?
What is meant by pixel loss?
What are dimensions? ae_dims ed_dims?
"figure out the max dimensions that you can use given your GPU's Memory" how?
What is "NN"?
What is "style power"?

Taking this one as an example:



I don't see any difference, what makes SAE better?

Sorry for the tons of Questions, jus trying to learn, improve and understand this Technology.

Thanks to anyone for answers and explanations.


@iperov if you could shortly explain, it would help me understand this technology way better and help me improve my fakes.
 

wbdragon

DF Pleb
halo2k said:
@iperov if you could shortly explain, it would help me understand this technology way better and help me improve my fakes.

I would like to suggest a glossary for terms that are used in the DFL guide.

 I have read it many times, and have gone through many post and searches to try to understand the terms. Sometimes through my experiments and trail and error, I am able to deduce what they mean. But I doubt myself.
 

halo2k

DF Pleb
Verified Video Creator
Well i think it is on his behalf that we understand and learn from this software. Of course i don't use something, i don't even know what it does, especially not in this scenario, where a wrong setting can destroy days over days of training.
 
Top