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.

Daisy Ridley HQ Faceset

mondomonger

DF Admirer
Verified Video Creator
Per my Faceset request for high quality celebrity facesets, enclosed is an example using Daisy Ridley:


New (Sept 9th) working link:     https://mega.nz/#!eKAjnCCJ!nmQvOTE-ENo14MBOPJrWpsI_Lk28f2r8SHiPAEfECKk


This faceset was created from 3  1080p movies:

Star Wars a New Hope
Star Wars the Last Jedi
Murder on the Orient Express

The faceset was edited prior to posting to get rid of blurry photos, photos with hands over face, hair over face, bad lighting ,etc. 

I used VirtualDub 2.0 to capture the frames with Daisy.  I then used Deepfacelab to extract the faces.     (You can use Fakeapp to do this via CPU --- it's slow, but it works well).        8 hours of training starting from Natalie->Lana created a new Daisy model that allowed me to make this within 11 hours:  https://mrdeepfakes.com/video/3047/daisy-ridley-dirty-talk-pov

You have a celebrity you want to see Deepfaked (Bollywood, Hollywood, 50's/60's/70's/80's, etc)  then post a very high quality faceset and the chances of a creator here picking up that faceset and running with it increases dramatically.      Give a faceset of similar quality to what I have posted, pick one of my clips, and I'll be happy to make something for you.
 

Cerciusx

DF Pleb
Not to be picky but there hardly any variance of expression here, and they face-forward shots all look essentially identical. is this a set you use?
 

mondomonger

DF Admirer
Verified Video Creator
Cerciusx -

"is this the set you use?" Seriously? You think I would post a faceset I don't use? To answer your question, yes, I used this faceset on the Daisy Ridley vid I posted today. 8 hours of training from a mature Natalie Portman-> Lana model turned this Daisy faceset into the "Dirty Talk POV" video I posted today.

I get better results when I capture at full FPS and end up with 3,000-5,000 faces. I generally keep all that have 2 eyes showing and are clear. I'm not sure, but I think varied lighting also can be helpful. If you want to see the difference, see my older Fakeapp version of this same Daisy Ridley Dirty Talk POV clip. You'll notice that the 30fps version properly talks and makes expressions while the older H64 model (with maybe 1,000 frames the faceset) is not as expressive.

Yes, you end up with a lot of face forward, but those extra frames help to properly teach your model to talk, help the mouth up to maximum, give you shots of the inside of the mouth which is important with H128 models, etc.

I've pretty much given up on profiles with DFL. They suck, so I generally don't include in the faceset. My general rule is: If you can't recognize the celeb in the Aligned folder based on face position (profile, nose up, nose far down), then don't include it in the faceset.
 

Cerciusx

DF Pleb
mondomonger said:
Cerciusx -

"is this the set you use?"     Seriously?    You think I would post a faceset I don't use?     To answer your question, yes, I used this faceset on the Daisy Ridley vid I posted today.   8 hours of training from a mature Natalie Portman-> Lana model turned this Daisy faceset into the "Dirty Talk POV" video I posted today.

I get better results when I capture at full FPS and end up with 3,000-5,000 faces.    I generally keep all that have 2 eyes showing and are clear.    I'm not sure, but I think varied lighting also can be helpful.     If you want to see the difference, see my older Fakeapp version of this same Daisy Ridley Dirty Talk POV clip.   You'll notice that the 30fps version properly talks and makes expressions while the older H64 model (with maybe 1,000 frames the faceset) is not as expressive.

Yes, you end up with a lot of face forward, but those extra frames help to properly teach your model to talk, help the mouth up to maximum,  give you shots of the inside of the mouth which is important with H128 models, etc.

I've pretty much given up on profiles with DFL.    They suck, so I generally don't include in the faceset.     My general rule is:   If you can't recognize the celeb in the Aligned folder based on face position (profile, nose up, nose far down), then don't include it in the faceset.

Ah ok man. I just didnt see a lot of variance so I didn't want to train on it if it didn't have a know good result. No bad blood man.
 

mondomonger

DF Admirer
Verified Video Creator
No worries man. You are probably right in that the faceset could use further editing. There are scenes where Daisy doesn't move and I've captured at 30fps, so that could be cleaned up. My overall point is that I have had considerably better results with high FPS captures (20-30) and not sweating the small differences for each face.

As you know, a good bit of my training is on a Geforce 1060. With DFL, the batch size is only 12, but I can still get results like you see here with 8 hours of training on an existing mature model.

This video: https://mrdeepfakes.com/video/3047/daisy-ridley-dirty-talk-pov

Started out as a Natalie Portman -> Lana model. I think Natalie and Daisy have similar facial features, so I recycled that model using the faceset I posted. 8 hours of training to Lana Rhoades master faceset, then about 3 hours of training to this specific clip = the result you see. It's far from perfect, but I am finding overall that Movie Lighting is far superior to Youtube Interview lighting in achieving good results.

I'm not a technical expert, so this is just based on my trial and error experience. My latest vids with: Kristen Bell, Charlize Theron, Emma Watson, Scarlett Johansson, and Evangeline Lily all come from either movie facesets or 1080p captures and all at a high FPS.


Last comment:

I had a big problem in past videos where the nose did not show up properly. Using Youtube interviews, the definition and color of the nose in the faceset was often hard to discern. With movies, this problem goes away. The lighting is varied and subtle enough to properly frame out the nose (even in nose down "going in for the blowjob!" positions.

All of my vids use H128 models and I do not use Pixel Correction as in my experience, it sucks.
 

greywolf22

DF Vagrant
Thank you so much! Please do STPeach face set when you can. There's also a thread about her set in requests section
 

mondomonger

DF Admirer
Verified Video Creator
Just to add to this, the CPU version of Deefacelab can make a faceset for you. No matter what your rig, you can download a movie, VirtualDub frames from it, then use Deepfacelab to extract to a proper faceset.

The "Mila Kunis Sucks Your Dick" was 100% Forgetting Sarah Marshall Virtualdub screencaps with movie to final faceset taking about 60-90 minutes of work. (Mila's getting a very big upgrade at the moment)

You want the celeb here on MrDeepfakes, then post the faceset and give me a shout.
 

deepfuck411

DF Vagrant
mondomonger said:
As you know, a good bit of my training is on a Geforce 1060.   With DFL, the batch size is only 12, but I can still get results like you see here with 8 hours of training on an existing mature model.

All of my vids use H128 models and I do not use Pixel Correction as in my experience, it sucks.

Hi,

I've a GTX 1060 3Go and I train my model using H64, batch size 4, light auto extracter.

TensorFlow crash without light_ae on and I'm not sure about the batch size.

H128 makes it crash too.

Not enough memory I think.

What about you ?
 

rain

DF Pleb
Verified Video Creator
Have the 100+ bj pics at the end been added by accident, or does that help the set? Thanks.
 

mondomonger

DF Admirer
Verified Video Creator
In my experience with H128, having dick in the mouth in the SRC helps to eliminate the lower set of teeth showing through on a blowjob conversion. Normally you see that half moon of white teeth through the dick. With dick training you can eliminate this. Maybe it helps for dick to lips clarify, but so far the difference in minimal.

Also a correction: With VirtualDub 2.0 I am capturing at the default framerate of 10 fps.
 
Top