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.

Mondomonger's Base H128 Model for Deepfacelab

mondomonger

DF Admirer
Verified Video Creator
This is an updated link.

H128 model, trained to 441K. Margot Robbie to Cherry Kiss. (Used through "Fiesty Fisting" video)

New LInk from Oct 22 2019: https://mega.nz/#!3TRHTCrK!ndUVr4cE_XthPJiV7uVrn0CbJET9bEoO5xlnXcuQAhI


Best conversion settings in my opinion are: 1) Overlay, 4) Fan_dst, Blur = 300 to 400, RCT Color Match. I leave all other settings at default. I sometimes adjust the Erode to -25 if I want more Celebrity face coverage.
 

666VR999

DF Enthusiast
Awesome! Thanks again, I got your FakeApp base model I'm playing with at the moment. I now supposedly got a system just good enough for GPU training (at least on FakeApp, GT 750m) so I can spend some time tweaking your fakeapp model to other celebs. I tried to use DFL but turns out GT750m computer capability is too low - however I'm going to see if I can use your new DFL model to again, "redirect" a base model, reverting to CPU if necessary (if possible!) Cheers!
 

LewiesMan

DF Pleb
Thanks Mondo! hope this works in CPU mode... my new GT-710 is "just" below minimum (it has 3.5 compute - you need 3.7)
 

XxXDarknessXxX

DF Vagrant
Thanks! How does using one model for multiple purposes works out for you? You have to use similar faces and lighting or just swap the dst and scr images and train away?
 

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
LewiesMan said:
Thanks Mondo! hope this works in CPU mode... my new GT-710 is "just" below minimum (it has 3.5 compute - you need 3.7)

Only H64 will work in CPU. This model will only work using GPU.
 

sotsog

DF Vagrant
Hello! Thanks for this.
I´m new to all of this and i have 3 questions, sorry, jeje:
1- I´m starting with just swaping girl´s faces in pictures, but i´d love to do videos in the future. But, in any of these cases, if i´m not using M. Robbie or C. kiss, but two other girls, and instead of starting from zero, i put THIS in "model" and  it´s going to save me, and everyone, hours and hours of training?
2- Or is it only when i do porn vids?
1- I f any of the pevious answers is yes, I downloaded 1 minute ago aStarter model from you (thanks again!) . Wich one is better for what i want (save training for swaping of all kind (from normal clothed girls pictures one day, perhaps porn another day, but not Margot Robbie)
THANKS A LOT AND GREAT JOB!!!
 

iperov

DF Enthusiast
Developer
I am announcing that all models except SAE will be removed in next DFL build, because SAE contains all models, but better, due to multiscale decoder + pixel loss. 
People continue to create crap fakes with old models based on only dssim loss, that produces less natural face and visible jitter, that make bad advertising for DFL.
DSSIM loss good only for first 15k epochs.
Multiscale decoder of SAE gaurantees fully recognized eyes. Pixel loss gaurantees natural face.

If you want to get classic model such as H128 or LIAEF, just set style values to zeros.
Look at my last fake created with last public build 
also new improvements to SAE are coming soon.
 

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
iperov said:
I am announcing that all models except SAE will be removed in next DFL build, because SAE contains all models, but better, due to multiscale decoder + pixel loss. 
People continue to create crap fakes with old models based on only dssim loss, that produces less natural face and visible jitter, that make bad advertising for DFL.
DSSIM loss good only for first 15k epochs.
Multiscale decoder of SAE gaurantees fully recognized eyes. Pixel loss gaurantees natural face.

If you want to get classic model such as H128 or LIAEF, just set style values to zeros.
Look at my last fake created with last public build 
also new improvements to SAE are coming soon.


I have a ton of old H128 models. Will they be able to convert to SAE? If not can you at least leave a download for older versions so that users can re-use old models? I have like 40+ H128 models.
 

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
sotsog said:
Hello! Thanks for this.
I´m new to all of this and i have 3 questions, sorry, jeje:
1- I´m starting with just swaping girl´s faces in pictures, but i´d love to do videos in the future. But, in any of these cases, if i´m not using M. Robbie or C. kiss, but two other girls, and instead of starting from zero, i put THIS in "model" and  it´s going to save me, and everyone, hours and hours of training?
2- Or is it only when i do porn vids?
1- I f any of the pevious answers is yes, I downloaded 1 minute ago aStarter model from you (thanks again!) . Wich one is better for what i want (save training for swaping of all kind (from normal clothed girls pictures one day, perhaps porn another day, but not Margot Robbie)
THANKS A LOT AND GREAT JOB!!!

If the models have a similar facial structure it might save you time. If they are really different it likely won't save any time.
 

mondomonger

DF Admirer
Verified Video Creator
Thank you for not removing old models! I too have about 20-30 trained H128 models.

That being said, I'll give SAE a try since you mention that it is the best.
 

mondomonger

DF Admirer
Verified Video Creator
iperov said:
i decided not to remove old models :D


I'm unable to get SAE model training to work on a 6GB Nvidia 1060 card.  I'm using all of the default options.   Sometimes it just won't get to the point of traing (just ends and says Press Any Key in the command line window).   Sometimes it will bring up the preview window for about 15 seconds, then stop working.

Is there a greater than 6GB memory requirement?
 

iperov

DF Enthusiast
Developer
mondomonger said:
iperov said:
i decided not to remove old models :D


I'm unable to get SAE model training to work on a 6GB Nvidia 1060 card.  I'm using all of the default options.   Sometimes it just won't get to the point of traing (just ends and says Press Any Key in the command line window).   Sometimes it will bring up the preview window for about 15 seconds, then stop working.

Is there a greater than 6GB memory requirement?

did you read help for every option in console?
 

mondomonger

DF Admirer
Verified Video Creator
D'oh! Adjusted the startup model options and was able to get SAE training started one 2 computers. I should be able to post results within 48 hours. Should I expect more natural color output on SAE trained models vs H128? I've had problems with the trained models not getting close to the DST face color with H128. If I turn histogram match on, the color matches, but the overall color of the face looks grey and fake looking - almost like it is drawn. Should I expect better results from SAE?


Very dumb question of the day: The batch size equals the amount of photos from SRC and DST used to compare? My 1060 w/ 6GB is running at a batch size of 6. Is that only 6 random photos compared on the SRC and DST side per epoch?
 

iperov

DF Enthusiast
Developer
mondomonger said:
Should I expect better results from SAE?
all depends on scene

mondomonger said:
 Is that only 6 random photos compared on the SRC and DST side per epoch?

yes.

My GTX1060 6GB runs with 16 batch size on SAE config df-512-42
 

viennaa89

DF Vagrant
Hey guys hope u dont mind I ask for some help ,for some reason my deepfacelab wont train.
When I click on "train H64" or any of the other training ones it'll just give me an endless amount of writing with no actual training having started. 
Ive tried closing all programs and starting again but I get the same thing, I dont know why this is happening when it was working perfectly not long ago?
 
D

deleted1

Guest
mondomonger said:
Enclosed is a link to my base model for Margot Robbie -> Cherry Kiss that was used as the initial starting point for the "Anal Queen" series of vids I posted recently.    This is trained to 250k-300k+ epochs,  Deepfacelab,   H128 model.  

https://mega.nz/#!SPYUAYyC!plKrXWvUiRp57suE5Lo8dFFDstTio3RAPcZrsi3ksns


In my experience, I can successfully train a new celeb to this model in 5 hours on a GTX1080. 

(I'm new to Mega, so let me know if the link does not work)

Thank you!! Your model is great. Now people are talking about SAE so can you share us a SAE trained model for testing purpose, please.
 

BatFaLl

DF Vagrant
mondomonger said:
Enclosed is a link to my base model for Margot Robbie -> Cherry Kiss that was used as the initial starting point for the "Anal Queen" series of vids I posted recently.    This is trained to 250k-300k+ epochs,  Deepfacelab,   H128 model.  

https://mega.nz/#!SPYUAYyC!plKrXWvUiRp57suE5Lo8dFFDstTio3RAPcZrsi3ksns


In my experience, I can successfully train a new celeb to this model in 5 hours on a GTX1080. 

(I'm new to Mega, so let me know if the link does not work)

Hey,

Thanks for providing this.

One question though: you said you can train a new celeb with this model in less than 5 hours. Does it mean that I can use a previously trained model and use it for a new video/faceset? How does it work? Do I need to keep the src and just swap the dst? Does it need to be the same person in both dst?
 
Top