MrDeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
Al_BundyHow to best train/convert frontal face BJ (with SAE)?
#1
I'm not very new to Deep Faking but I surely lack understanding of some things here and there. I thought that the best model we now should use is SAE (maybe some other models have their benefits in edge cases).

I have a BJ scene with frontal face, so the sausage :-) is blocking the lower lip and chin area (and below).

When I train with SAE and use the default settings during conversion I get a result with the src face overlapping the ara where the sausage was, so no magic seems to happen there and the obstruction ( I guess this is also an obstruction) is not in the resulting image.

I'm not sure how to solve this issue and reading the forum made me even more confused. Should I use the advanced mask editor and mask the sausage out in each frame? Then after doing this training the model again a lot so it knows the face and also the masked out area and then convert it using mode overlay (or seamless?) and mask mode "fan-dst"?

I trained it for 30k iterations without masking out the sausage and then masking it out in 30 frames. Then made maybe 1k iterations (only) with this and tried SAE debug convert with "FAN-DST" but what happens is, that not the complete sausage will get shown, some parts of the resulting (src) face will mix up with it. Maybe it is because I haven't trained it enough after masking it out, or do I need to mask more space out?

Hope anyone can push me in the right direction.
#2
Don't use any kind of masking.

Train normally with the obstructions in the training process.

When converting, use any FAN-X. FAN-X is already pre-trained for obstructions.
~ Fake it till you make it ~
You are not allowed to view links. Register or Login to view.
You are not allowed to view links. Register or Login to view.
#3
(06-08-2019, 06:03 PM)dpfks Wrote: You are not allowed to view links. Register or Login to view.Don't use any kind of masking.

Train normally with the obstructions in the training process.

When converting, use any FAN-X. FAN-X is already pre-trained for obstructions.

Okay, but there are 4 different FAN masking options thereSad3) FAN-prd, (4) FAN-dst , (5) FAN-prd*FAN-dst (6) learned*FAN-prd*FAN-dst

you say, it doesn't matter which one I take? Why do 4 different exist then?

I tried it now with option 4, discarding my masks, it looks not to bad but in some images the "sausage" gets mixed up a big with the src image - is there a way to  improve it?
#4
I realized that my dst face set contained only face pics with obstructions. Maybe this is the issue, I now also included faces without instruction of src face into my aligned images set for training. Is this the right approach and I guess I can leave the images with the obstructions in there (also for training)?
#5
(06-09-2019, 01:37 PM)Al_Bundy Wrote: You are not allowed to view links. Register or Login to view.I realized that my dst face set contained only face pics with obstructions. Maybe this is the issue, I now also included faces without instruction of src face into my aligned images set for training. Is this the right approach and I guess I can leave the images with the obstructions in there (also for training)?

Not sure of the answer to this question, just here to say that I had good results using (6) learned*FAN-prd*FAN-dst
I accept donations with Tokens, Bitcoin, Paypal or Credit Card

To request a paid deepfake, just send me a private message.

All my videos so far : You are not allowed to view links. Register or Login to view.

Consider donating to the creator of DFL :

You are not allowed to view links. Register or Login to view.
#6
(06-09-2019, 04:27 PM)SPT Wrote: You are not allowed to view links. Register or Login to view.
(06-09-2019, 01:37 PM)Al_Bundy Wrote: You are not allowed to view links. Register or Login to view.I realized that my dst face set contained only face pics with obstructions. Maybe this is the issue, I now also included faces without instruction of src face into my aligned images set for training. Is this the right approach and I guess I can leave the images with the obstructions in there (also for training)?

Not sure of the answer to this question, just here to say that I had good results using (6) learned*FAN-prd*FAN-dst

This mask option seems to be removed from the latest DFL version. Well at least I am training a new model with SAE, full face and masking learning. When I convert I cannot select this mask option (FAN-x). I only can choose between DST and "learned".
#7
Hi Al -

Hope you like my Kelly Bundy vids! For all of my videos, I leave dick in the mouth as long as it is dick only with no fingers. Best is lips locked on the dick, 2 eyes forward. This is never a problem for me in training. Nearly all of my vids are H128. SAE takes longer to train and about 4x longer to convert with Fan-dst turned on.

The angle of the BJ is important. See my "_______ sucks your dick" series to get a better idea.

I forgot to mention, be sure to also have the full face of the porn star showing for training. For my vids, I always include pornstar facing forward, talking, unobstructed face. Say I want to do a 30 sec blowjob vid that is only blowjob. I'll still include full face, foward, talking captures of the pornstar (maybe 1,000 pics) in the main base training DST set.
#8
(06-12-2019, 05:37 PM)Al_Bundy Wrote: You are not allowed to view links. Register or Login to view.This mask option seems to be removed from the latest DFL version. Well at least I am training a new model with SAE, full face and masking learning. When I convert I cannot select this mask option (FAN-x). I only can choose between DST and "learned".

Learned means it uses the mask that the model learned during training. So that's the one you want
#9
(06-12-2019, 05:37 PM)Al_Bundy Wrote: You are not allowed to view links. Register or Login to view.
(06-09-2019, 04:27 PM)SPT Wrote: You are not allowed to view links. Register or Login to view.
(06-09-2019, 01:37 PM)Al_Bundy Wrote: You are not allowed to view links. Register or Login to view.I realized that my dst face set contained only face pics with obstructions. Maybe this is the issue, I now also included faces without instruction of src face into my aligned images set for training. Is this the right approach and I guess I can leave the images with the obstructions in there (also for training)?

Not sure of the answer to this question, just here to say that I had good results using (6) learned*FAN-prd*FAN-dst

This mask option seems to be removed from the latest DFL version. Well at least I am training a new model with SAE, full face and masking learning. When I convert I cannot select this mask option (FAN-x). I only can choose between DST and "learned".

What version are you using... You should have FAN-X.
~ Fake it till you make it ~
You are not allowed to view links. Register or Login to view.
You are not allowed to view links. Register or Login to view.
#10
Quote:What version are you using... You should have FAN-X.

I guess the latest from Jun 2019. And I also have the option now, don't think I have changed anything only that I trained more iterations. Maybe it is not there when you have only trained some dozen iterations...

Forum Jump:

Users browsing this thread: 1 Guest(s)