MrDeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
qgggqCollab SAE settings
#1
Ok guys,

i've been trying to do some deepfake for few weeks now, works great in h128 etc, somehow know how to have some quite good results in h128 but i want to switch to SAE as it seems to be just better with face details.

i've tried few things , starting with a low batch , low "details" settings ( no clipgrad, optimization at 1 etc ) but it takes forever to get to 30k even on a T4. ( lol @ k80 which seems to be 2x slower )

i've tried to start directly at batch 24 , optimization 3 , clipgrad , face & background style power at 10 etc like in the guide , but after 7h , details were there, but it was quite shitty , eyes were not alligned at all , sometimes it looked like a painting with weird angles etc


What am i missing ? i'm working on 6k+ alligned faces on both sides , with differents angles and facial expressions... but idk how to have a perfect result with sae.
#2
Anyone have some tips how to improve fakes using SAE? I have already done some pretty good videos but in all of them quality of faces are not so good. Here is example of my training:

[Image: 33Z8tO6h.jpg]

I do not know how to improve it anymore. I tried pixel lost but it always collapse after few hundred iterations...

Here is example what others can do with this model... I just... wow...

[Image: IwZ6NIY.jpg]
#3
(07-28-2019, 11:55 AM)Cabernet Wrote: You are not allowed to view links. Register or Login to view.Anyone have some tips how to improve fakes using SAE? I have already done some pretty good videos but in all of them quality of faces are not so good. Here is example of my training:

[Image: 33Z8tO6h.jpg]

I do not know how to improve it anymore. I tried pixel lost but it always collapse after few hundred iterations...

Here is example what others can do with this model... I just... wow...

[Image: IwZ6NIY.jpg]

The first biggest problem you have is you  are comparing your model with ctrl shift one !

That guy literally spends months I guess to get 1080p facesets  from all the fuking angles  from the movies you have never even heard of and  whatever limit of which you can imagine.
When you put soo much effort in that the fruits of it will show . I guess its his only job now . Heard he is GRAPHIC ILLUSTRATOR.

He must be also doing post processing which I believe you won't do.

After seeing your result I guess You are making errors in choosing the wrong model settings and not enough/high quality facesets at first place .

Did you pretrain your model ?

Now see I am also training my model and I am getting this result in just 80k iterations 

You are not allowed to view links. Register or Login to view.

[Image: open?id=1iNooQQRQ8z6IdKE4ygUSZFgy0j8mpHSK]

Work on what you feed to the DFL . The higher quality you put the better you get . Also pretrain it
#4
Almost more important than having the right settings is having good face sets. Are all your pictures very high resolution to begin with, all blurry images manually removed, all angles covered etc? I generally have to train to about 40-50k before I start seeing proper usable models with SAE, where details are starting to come through etc. You can try to run the "data_src final" and "data_dst final" task to reduce the number of faces while still maintaining all angles available, which might improve / speed up the initial training. You might want to add the resulting "trash" folder back in when you get past 60-70k interations. I have had some good experience with this at least.

I typically run with about 3000 faces in each set (filtered with the final task) and use a batch size of 17. When the loss starts to flatten out, I enable rct color transfer and add the trash images back in to tweak the model for the remaining details. Been using Colab for all my training lately, and T4 is the fastest GPU I have been getting. K80 takes about twice as long per iteration, yeah.

Forum Jump:

Users browsing this thread: 1 Guest(s)