Mr DeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
  •  Previous
  • 1
  • ...
  • 22
  • 23
  • 24(current)
  • 25
  • 26
  • ...
  • 123
  • Next 
dpfksDeepFaceLab Explained and Usage Tutorial
(01-07-2019, 09:33 PM)Kenturin Wrote: You are not allowed to view links. Register or Login to view.
(01-07-2019, 09:23 PM)iperov Wrote: You are not allowed to view links. Register or Login to view.GeForce GTX 780 compute capability 3.5, but minimum required 3.7 for tensorflow 1.11.0.
You are not allowed to view links. Register or Login to view.
so your card not working with dfl

Ah thanks for your time, I overlooked that obvious bit.  Even so, I've been impressed with even the CPU results on the H64 models in DFL!  All the more reason to upgrade to a non-stone age GPU!

you can downgrade to tf 1.10.0 which works with 3.5 cap

run DeepFaceLab\_internal\bin\commandline.cmd

python -m pip uninstall tensorflow-gpu
python -m pip install tensorflow-gpu==1.10.0
(01-07-2019, 09:48 PM)iperov Wrote: You are not allowed to view links. Register or Login to view.
(01-07-2019, 09:33 PM)Kenturin Wrote: You are not allowed to view links. Register or Login to view.
(01-07-2019, 09:23 PM)iperov Wrote: You are not allowed to view links. Register or Login to view.GeForce GTX 780 compute capability 3.5, but minimum required 3.7 for tensorflow 1.11.0.
You are not allowed to view links. Register or Login to view.
so your card not working with dfl

Ah thanks for your time, I overlooked that obvious bit.  Even so, I've been impressed with even the CPU results on the H64 models in DFL!  All the more reason to upgrade to a non-stone age GPU!

you can downgrade to tf 1.10.0 which works with 3.5 cap

run DeepFaceLab\_internal\bin\commandline.cmd

python -m pip uninstall tensorflow-gpu
python -m pip install tensorflow-gpu==1.10.0

Worked like a charm!  Now seeing 88% vRAM load and epoch time dropped from 4000ms to 240ms, lol.  Thanks for all of the help!
@iperov @dpfks have a question for you guyz what is the real difference between DF Training and H128 Training ?
Can you please guyz tell me why one is better than the others etc .. ?
(01-08-2019, 01:03 AM)Akanee Wrote: You are not allowed to view links. Register or Login to view.@iperov @dpfks have a question for you guyz what is the real difference between DF Training and H128 Training ?
Can you please guyz tell me why one is better than the others etc .. ?

Literally half a dozen posts ago...

(01-07-2019, 06:23 PM)iperov Wrote: You are not allowed to view links. Register or Login to view.Every model good for specific scenes and faces.
H64 - good for straight faces as demo and for low vram.
H128 - good for straight faces, gives highest resolution and details possible in 2019. Absolute best for asian faces, because they are flat, similar and evenly lighted with clear skin.
DF - good for side faces in drawback of resolution and details, covers more area of cheeks. Keeps face unmorphed. Good for similar face shapes.
LIAE - can partially fix dissimilar faces shape in drawback of less recognizable face.
SAE - no matter how similar faces, src face will be morphed onto dst face, can make face absolutely unrecognizable. Model can be collapsed on some scenes. Easy to overlay final face because dst background also predicted.


watch basic DFL workflow You are not allowed to view links. Register or Login to view.
(01-07-2019, 06:23 PM)iperov Wrote: You are not allowed to view links. Register or Login to view.
(01-07-2019, 05:16 PM)halo2k Wrote: You are not allowed to view links. Register or Login to view. i don't want to waste 1 week of Training to find out, it was the wrong model for my Video.

unfortunately deepfaking is time/eletricpower consuming topic and has a lot of nuances.

Every model good for specific scenes and faces.
H64 - good for straight faces as demo and for low vram.
H128 - good for straight faces, gives highest resolution and details possible in 2019. Absolute best for asian faces, because they are flat, similar and evenly lighted with clear skin.
DF - good for side faces in drawback of resolution and details, covers more area of cheeks. Keeps face unmorphed. Good for similar face shapes.
LIAE - can partially fix dissimilar faces shape in drawback of less recognizable face.
SAE - no matter how similar faces, src face will be morphed onto dst face, can make face absolutely unrecognizable. Model can be collapsed on some scenes. Easy to overlay final face because dst background also predicted.

@iperov So if I am understanding this correctly:
H128 will attempt to place the src face over the dst face
SAE will attempt to morph the src face into the dst face

Or vice versa? or something completely different?
Can you add more files to src\aligned later on? With the old apps you had an aligned.json which kind of confused the issue, but now you only have the PNG's and you can add/remove files as you wish?

Also, is there an ideal number of aligned images (for 128 and 64 variants) for an 11GB video card? or the more the merrier?

Thanks so much iperov.
Regarding a few previous questions concerning required training time, I crunched some numbers using the newer SAE model.

Using an RTX 2080 (yes I upgraded from my steam engine-era GTX 780  Tongue and I did reinstall tensorflow 1.11.0 )
Resolution: 128
AE_dims: 256
Full face
batch size: 16

I've been looking at an average time of 900ms/epoch

Assuming I didn't fat finger the math, that's about 4k epochs/hour or about 25 hours to reach 100k epochs.

This obviously says nothing in regards to the quality of the final model at the 100k point.  That of course depends heavily on the quality of the data sets as well as conversion options.
Just wanted to point out that you can get a feel for how many "epochs per hour" you can expect from your card at various settings in each model with a little math.

On a side note, my data set is shit, so I look forward to whatever Lovecraftian monstrosity this round of training creates.
(01-08-2019, 09:49 AM)frosty3907 Wrote: You are not allowed to view links. Register or Login to view.Can you add more files to src\aligned later on? With the old apps you had an aligned.json which kind of confused the issue, but now you only have the PNG's and you can add/remove files as you wish?

Also, is there an ideal number of aligned images (for 128 and 64 variants) for an 11GB video card? or the more the merrier?

Thanks so much iperov.

Yes, you can add and remove from the data_src\aligned as you please. But you cannot do this for the data_dst unless they are from the same extraction process. For example, I would usually train with less data_dst by removing some images are are blurry, duplicates, or semi-obstructed. Then for the final conversion, I add all the dst images back.

There is no ideal number of images for facesets. But the less is better; however you need to make sure you cover all your different angles. Typically 1000-4000 images are decent.
I am getting a fairly distinct white line across the tongue when using Deepfacelab. What settings should I use to minimize this?

When using Fakeapp 2.2, at the correct angle, when the tongue sticks out at about a 3/4 angle, it lines up with the replaced face and can look very realistic. The exact same clip with Deepfacelab ends up with a relatively thick, distinct whitish line across the tongue. I'm using the overlay option, match color option when converting, but not changing the blur setting. How to eliminate that white line across the tongue?

I've upgraded HD space and am making the move to H128, but I don't want to give up the cum on tongue shots!
(01-09-2019, 03:23 AM)mondomonger Wrote: You are not allowed to view links. Register or Login to view.I am getting a fairly distinct white line across the tongue when using Deepfacelab.  What settings should I use to minimize this?

When using Fakeapp 2.2, at the correct angle, when the tongue sticks out at about a 3/4 angle, it lines up with the replaced face and can look very realistic.    The exact same clip with Deepfacelab ends up with a relatively thick, distinct whitish line across the tongue.     I'm using the overlay option, match color option when converting, but not changing the blur setting.  How to eliminate that white line across the tongue?

I've upgraded HD space and am making the move to H128, but I don't want to give up the cum on tongue shots!

Try playing around with the erode and blur settings. When I swapped over I had to keep testing to see which values gave the best results to prevent that boxed look. This depends of the type of scene.
  •  Previous
  • 1
  • ...
  • 22
  • 23
  • 24(current)
  • 25
  • 26
  • ...
  • 123
  • Next 

Forum Jump:

Users browsing this thread: asd, oskmartin3, 4 Guest(s)