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.

Pretraining NSFW faceset only women, celebs + pornstars [Full face]

SPT

Moderator
Staff member
Moderator
Verified Video Creator
This works only for pretraining !

It's made of celebs from CelebA pack, removed all men, removed 2000-3000 shittiest quality pictures, added 2000-3000 porn faces.

Note : Works only for full face, because porn face in this one are extracted in full face.

If you want one working with both Full Face and Whole Face, choose this new version : https://mrdeepfakes.com/forums/thre...ly-women-celebs-pornstars-wholeface-full-face


Usage :

Download, put it in DeepFaceLab_NVIDIA\_internal\pretrain_CelebA  (save the original pak if needed)

Choose a new model, any archi and resolution you want, don't use any options besides Pre-train, and random warp until 50k-100k training, then disable RW. Train up to 175-200k (can go up to 400k or more if you feel like it, but it's usable from 175k).

Save this model somewhere, use a copy of it to start training a new deepfake, by disabling Pretrain parameter. It will start from 0 but will go up to a satisfying level faster (with 175k pretraining, 60k was enough training for the 1st deepfake I did with it). Re-use copy of base pretrained model for each new celeb you'll fake.

More infos and precisions about pretraining on the guide by Tutsmybarreh.
https://mrdeepfakes.com/forums/thread-guide-deepfacelab-2-0-explained-and-tutorials-recommended

Download :
https://mega.nz/file/sp1DxaQQ#bm39vgSnL3tEISex29qmuYFzuMNoDkN-eQMQCS8tqD8


PS : let me know if it is the right place to post this.
 

1D0F4K35

DF Enthusiast
Verified Video Creator
This is a great post. Thanks for this. Just the job :)

Mind if I ask some questions I hope you may be able to help me with.

If I set this up on COLAB then download the SAEHD model to my PC will my DFL v1 be able to read it?
I have to use DFL v1 because I have an AMD GPU.

I use V1 COLAB too but the models I make on there cant be read by my PC DFL v1 for any additional training. Any ideas why that would be?
I should say, they are OK to convert from but not to continue training with. Does all this make sense?
 

SPT

Moderator
Staff member
Moderator
Verified Video Creator
bobbydavro said:
This is a great post. Thanks for this. Just the job :)

Mind if I ask some questions I hope you may be able to help me with.

If I set this up on COLAB then download the SAEHD model to my PC will my DFL v1 be able to read it?
I have to use DFL v1 because I have an AMD GPU.

I use V1 COLAB too but the models I make on there cant be read by my PC DFL v1 for any additional training. Any ideas why that would be?
I should say, they are OK to convert from but not to continue training with. Does all this make sense?

Never tried Colab, wether with DFL 1 or 2. But if you use DFL colab 1 and DFL 1, you seem right.

This pretraining thing is something I didn't do with DFL 1.0, but I just looked into my DFL 1.0 folder and it seems there is the same folder structure to use this .pak

So this should work with DFL 1.0 but I didn't try.

How it works is you train it (either on your comp or collab) (Btw I don't know where's the pre-train setting for DFL 1.0, you'll have to check the DFL 1 guide for that.)

There's no src/dst it just trains everything mixed in the .pak

Then when it's trained enough keep a copy of the model somewhere, and use another copy as base to train your real video with src and dst as usual.

In DFL 2 when you click train it will show the model's name, then you just have to change the pre-train option to "no" and it will start from 0 but it will be faster to train your usual src/dst fake because of the pre training.

Again look into the DFL 1.0 guide.

Also, I'm not sure but maybe the fact that Collab uses NVidia cards and you use AMD at home can cause conflicts. You'll have to ask more knwoledgeable persons about this.

Hope it's a bit clearer, good luck.
 

1D0F4K35

DF Enthusiast
Verified Video Creator
SPT said:
Hope it's a bit clearer, good luck.

It is and thanks a lot for that.
I've got it training (on my PC for now) and it's working fine.
I was just hoping to be able to create it on COLAB for the sake of speed and convenience, and then once built, use it on my PC but it seems the models I make on COLAB are never going to be interchangable with my PC for training.
Puzzling how there's no NVDIA/AMD conflict with the converting though. I'll ask around some more about that.
 

TMBDF

Moderator | Deepfake Creator | Guide maintainer
Staff member
Moderator
Verified Video Creator
Not sure how I've missed this thread, anyway, sticked it so it's more visible.
 

neurojunkie

DF Vagrant
Thanks a ton for this! Great work thanks for sharing.

Super random question but for some reason I had selected "random warp" and not random flip faces and I am now approaching 175K its. This turned into a question that I think might be interesting to have people far more knowledgeable than myself perhaps weigh in on.. What impact might 174K its of only random warp have on this pre-training process/pre-training in general?
(I get the point of flip faces to have a more versatile pre-train to prepare for the widest variety of angles in source material [so this is one way my pre-train is impacted negatively in my oversight from your directions]; so my question is really just regarding the overall hypothetical difference in having used random warp instead in this case on the quality/characteristics of the pre-train)
 

SPT

Moderator
Staff member
Moderator
Verified Video Creator
neurojunkie said:
Thanks a ton for this! Great work thanks for sharing.

Super random question but for some reason I had selected "random warp" and not random flip faces and I am now approaching 175K its. This turned into a question that I think might be interesting to have people far more knowledgeable than myself perhaps weigh in on.. What impact might 174K its of only random warp have on this pre-training process/pre-training in general?
(I get the point of flip faces to have a more versatile pre-train to prepare for the widest variety of angles in source material [so this is one way my pre-train is impacted negatively in my oversight from your directions]; so my question is really just regarding the overall hypothetical difference in having used random warp instead in this case on the quality/characteristics of the pre-train)

Sorry for an unprecise answer but : What I do is I stop RW once I feel loss values are progressing too slow. I just learnt yesterday that there's another school of thought that believes in using only RW during the whole training of a pretrain model and only deactivate RW for when you use it for a real src-dst model. And btw, I also don't know if it's better to reactivate RW for a bit once in src-dst stage or not. Or maybe it's a good idea only if you trained the pre-train model entirely in RW.

You need the advice from others here as I don't know definitive answers on that one.
 

neurojunkie

DF Vagrant
SPT said:
neurojunkie said:
Thanks a ton for this! Great work thanks for sharing.

Super random question but for some reason I had selected "random warp" and not random flip faces and I am now approaching 175K its. This turned into a question that I think might be interesting to have people far more knowledgeable than myself perhaps weigh in on.. What impact might 174K its of only random warp have on this pre-training process/pre-training in general?
(I get the point of flip faces to have a more versatile pre-train to prepare for the widest variety of angles in source material [so this is one way my pre-train is impacted negatively in my oversight from your directions]; so my question is really just regarding the overall hypothetical difference in having used random warp instead in this case on the quality/characteristics of the pre-train)

Sorry for an unprecise answer but : What I do is I stop RW once I feel loss values are progressing too slow. I just learnt yesterday that there's another school of thought that believes in using only RW during the whole training of a pretrain model and only deactivate RW for when you use it for a real src-dst model. And btw, I also don't know if it's better to reactivate RW for a bit once in src-dst stage or not. Or maybe it's a good idea only if you trained the pre-train model entirely in RW.

You need the advice from others here as I don't know definitive answers on that one.

Thanks for your response. Every bit of insight helps!
 

SPT

Moderator
Staff member
Moderator
Verified Video Creator
fjqm64dlxc said:
Forgive my ignorance but what is the advantage in using an NSFW faceset in relation to SFW?

I don't know, where did you get this idea ? 
Anyways, this set is quite old now, pretty much everyone uses wf facesets with 512+ res.
 

fjqm64dlxc

DF Vagrant
SPT said:
fjqm64dlxc said:
Forgive my ignorance but what is the advantage in using an NSFW faceset in relation to SFW?

I don't know, where did you get this idea ? 
Anyways, this set is quite old now, pretty much everyone uses wf facesets with 512+ res.

I just assume that no one will spend hours of life making a faceset with zero purpose, there has to be some advantage, and I was wondering what it is.
 

SPT

Moderator
Staff member
Moderator
Verified Video Creator
fjqm64dlxc said:
SPT said:
fjqm64dlxc said:
Forgive my ignorance but what is the advantage in using an NSFW faceset in relation to SFW?

I don't know, where did you get this idea ? 
Anyways, this set is quite old now, pretty much everyone uses wf facesets with 512+ res.

I just assume that no one will spend hours of life making a faceset with zero purpose, there has to be some advantage, and I was wondering what it is.

It was made to help with NSFW deepfakes, for SFW ones, I guess it doesn't hurt but it does nothing special.
 
Top