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.

Please explain what "pretraining a model" is and if it would be useful in my case

andax

DF Vagrant
My example:

I'm using footages of Hitler for src.

As for dst I have a lovely politican who would be thrilled by sharing more than thoughts with him.

This politician obviously has a lot of footage available.

My goal is to apply the src to different dst sets without changing the people.

Task A) src (Hitler) -> dst (Politician)
This is ok, it works perfectly. This guy really shares a lot with src.

Task B) use same src with different dst (same person, similar faces but still it's a new set)

If I replace only the dst images before starting Task B would this be considered pretraining?


If I'm complete mistaken here which is likely, is there a way to make future swaps (same people same angles) easier/faster by using data from earlier transformations?

Thank you. Please feel free to move my post to the appropriate place if this is not it.
 

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
For you since you'll be reusing the models over and over I don't think you need to use the pretrain function.

The pretrain function will train your model with many different faces and at different lighting and angles. This will ultimately make your final product look less like data_src
 

iperov

DF Enthusiast
Developer
@dpfks It seems you doing all to decrease rating of DFL.

RTFM:

Pretrain the model with large amount of various faces. This technique may help to train the fake with overly different face shapes and light conditions of src/dst data. Face will be look more like a morphed. To reduce the morph effect, some model files will be initialized but not be updated after pretrain: LIAE: inter_AB.h5 DF: encoder.h5. The longer you pretrain the model the more morphed face will look. After that, save and run the training again.

This fake trained on 3k iters pretrained model:



even little pretrain enough to recognize faces better
 

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
@"iperov" not really.

You showed everyone a pre-train vs . no pre-train example prior to you implementing this option. The result in my opinion are bad because the face is too morphed (yes I know that example pre-trained for a long time).

What I don't have examples of is lower pre-train 3-5k pretrain vs not.

This user is asking about re-using the model for multiple scenes. As he re-uses the model more the lighting conditions will improve...
 

iperov

DF Enthusiast
Developer
dpfks said:
The result in my opinion are bad because the face is too morphed (yes I know that example pre-trained for a long time).

you are wrong.
Result is the best, because there is NO RESULT at all without pretrain.
 

mechelina

DF Pleb
iperov said:
dpfks said:
The result in my opinion are bad because the face is too morphed (yes I know that example pre-trained for a long time).

you are wrong.
Result is the best, because there is NO RESULT at all without pretrain.

Personally I get stunning results with pre-training with different faces, so I agree with Iperov. I get faster results and the dst seems better trained in recognizing expressions in faces. It seems you are learning the AI first what faces are, how they are build and how they can move and act. Which seems to speed the proces. 

And yes, I am sure you will get results too where the face doesn't look like the celeb, but I personally think that you will have to look at the source and destiny to actually match. The more similar the source and destiny are in bone structure and the way they use their expressions in the face, the more the result will look like the celeb you choose as source.

If it was up to me, I would love a "pre"-pre-trained start file so the AI already has the knowledge of faces in general.
 
Top