MrDeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
  •  Previous
  • 1
  • ...
  • 106
  • 107
  • 108(current)
  • 109
  • 110
  • ...
  • 128
  • Next 
DeepFaceLab Explained and Usage Tutorial
(05-16-2019, 10:34 AM)DaddyLee Wrote: You are not allowed to view links. Register or Login to view.So, when I use an old trained model to start a new deepfake, it starts off by saying it's at the same iteration it was on before. Can it not start from zero? Am I eventually gonna have to start training with super high iterations being displayed and just try to keep track of the math?

just delete SAE or H128 files (depend wich mode you used for training)  in the model folder.
And start training all over.
(05-16-2019, 11:31 AM)Akanee Wrote: You are not allowed to view links. Register or Login to view.
(05-16-2019, 10:34 AM)DaddyLee Wrote: You are not allowed to view links. Register or Login to view.So, when I use an old trained model to start a new deepfake, it starts off by saying it's at the same iteration it was on before. Can it not start from zero? Am I eventually gonna have to start training with super high iterations being displayed and just try to keep track of the math?

just delete SAE or H128 files (depend wich mode you used for training)  in the model folder.
And start training all over.

Thanks so much! Will do that !

(05-16-2019, 12:30 PM)DaddyLee Wrote: You are not allowed to view links. Register or Login to view.
(05-16-2019, 11:31 AM)Akanee Wrote: You are not allowed to view links. Register or Login to view.
(05-16-2019, 10:34 AM)DaddyLee Wrote: You are not allowed to view links. Register or Login to view.So, when I use an old trained model to start a new deepfake, it starts off by saying it's at the same iteration it was on before. Can it not start from zero? Am I eventually gonna have to start training with super high iterations being displayed and just try to keep track of the math?

just delete SAE or H128 files (depend wich mode you used for training)  in the model folder.
And start training all over.

Thanks so much! Will do that !
Is it the 'SAE.data.dat file?

Because I just copied over the entire model folder from the workspace so I'm not sure which files in there to delete to do this correctly and have it start iterations from the beginning.
(05-16-2019, 01:20 AM)dpfks Wrote: You are not allowed to view links. Register or Login to view.CA weights - different method for model to calculate how different data_src vs data_dst is. No need to use.
wtf? where you read that?

official help for CA weights:

Initialize network with 'Convolution Aware' weights. This may help to achieve a higher accuracy model, but consumes a time at first run.
(05-16-2019, 12:30 PM)DaddyLee Wrote: You are not allowed to view links. Register or Login to view.
(05-16-2019, 11:31 AM)Akanee Wrote: You are not allowed to view links. Register or Login to view.
(05-16-2019, 10:34 AM)DaddyLee Wrote: You are not allowed to view links. Register or Login to view.So, when I use an old trained model to start a new deepfake, it starts off by saying it's at the same iteration it was on before. Can it not start from zero? Am I eventually gonna have to start training with super high iterations being displayed and just try to keep track of the math?

just delete SAE or H128 files (depend wich mode you used for training)  in the model folder.
And start training all over.

Thanks so much! Will do that !

(05-16-2019, 12:30 PM)DaddyLee Wrote: You are not allowed to view links. Register or Login to view.
(05-16-2019, 11:31 AM)Akanee Wrote: You are not allowed to view links. Register or Login to view.
(05-16-2019, 10:34 AM)DaddyLee Wrote: You are not allowed to view links. Register or Login to view.So, when I use an old trained model to start a new deepfake, it starts off by saying it's at the same iteration it was on before. Can it not start from zero? Am I eventually gonna have to start training with super high iterations being displayed and just try to keep track of the math?

just delete SAE or H128 files (depend wich mode you used for training)  in the model folder.
And start training all over.

Thanks so much! Will do that !
Is it the 'SAE.data.dat file?

Because I just copied over the entire model folder from the workspace so I'm not sure which files in there to delete to do this correctly and have it start iterations from the beginning.

No you can't restart the iterations.

Also, who cares about the math. Use your preview.

(05-16-2019, 06:00 PM)iperov Wrote: You are not allowed to view links. Register or Login to view.
(05-16-2019, 01:20 AM)dpfks Wrote: You are not allowed to view links. Register or Login to view.CA weights - different method for model to calculate how different data_src vs data_dst is. No need to use.
wtf? where you read that?

official help for CA weights:

Initialize network with 'Convolution Aware' weights. This may help to achieve a higher accuracy model, but consumes a time at first run.

My bad, will edit.
~ Fake it till you make it ~
You are not allowed to view links. Register or Login to view.
Thanks dpfaks!
Good Advice!

How does DFL deal with tongues?? Like when they stick their tongue out...

Anyone have experience with this? I guess I will find out eventually, lol, since there's a lot of it in my video...

DUDE!

Anyone who has a model trained to 100k... RE-USE IT

I CAN'T BELIEVE how much time it takes off! WOW. It's getting accurate results LIGHTNING quick!
When converting, and it says ' o faces found for 00516.png, copying without faces ' for example,

Is there some way to avoid this? Or is this just natural since you've deleted some misalignment etc? Or does it just mean that faces weren't detected in those frames?
DaddyLee

that comes from deleted alinged pics. dont worry about if u just deleted the crap :)
DaddyLee

I'm probably being captain obvious here. That message could be one of 2 things, first there are no faces found in that frame and the original frame is going to be copied, which is perfectly fine -or- second, if there is a face in that frame it's been deleted, possibly because you deleted it when cleaning up a faceset for training.

For my workflow, when I have a scene I create a few directories:

one is for the original set of aligned faces extracted which I clean up by removing anything that is not the face I want swapped. I keep every face of the pornstar even if blurry. This could be 10,000+ faces depending on how long the scene is, right? Images in this directory determine which face/frame will be swapped when you convert.

second directory I create is the faceset for training. These are a subset of the first, extracted images where I remove anything blurry and focus on trying to get a variety of expressions and angles. This is where you do all your sorts and trim out the crap. I shoot for 2-4,000 images here - best quality.  Because crap in = crap out.

You're not going to break anything if you modify the .bat files to extract from one folder and train from another. They are pretty self explanatory. 
Remove the ones you don't use and copy/edit the ones you make changes to. I have multiple convert scripts at any given time.

Not related, but 2 other tips I'll swear by:

Back your model up. often. Can be as simple as copying and renaming the directory.

Also, when converting, I take a subset of about 300-500 images from the full scene, throw then in a scene_test folder and convert those as a test. Only takes a few minutes and really lets me know how the model training is going.
(05-17-2019, 09:24 PM)feepdakenoob Wrote: You are not allowed to view links. Register or Login to view.DaddyLee

I'm probably being captain obvious here. That message could be one of 2 things, first there are no faces found in that frame and the original frame is going to be copied, which is perfectly fine -or- second, if there is a face in that frame it's been deleted, possibly because you deleted it when cleaning up a faceset for training.

For my workflow, when I have a scene I create a few directories:

one is for the original set of aligned faces extracted which I clean up by removing anything that is not the face I want swapped. I keep every face of the pornstar even if blurry. This could be 10,000+ faces depending on how long the scene is, right? Images in this directory determine which face/frame will be swapped when you convert.

second directory I create is the faceset for training. These are a subset of the first, extracted images where I remove anything blurry and focus on trying to get a variety of expressions and angles. This is where you do all your sorts and trim out the crap. I shoot for 2-4,000 images here - best quality.  Because crap in = crap out.

You're not going to break anything if you modify the .bat files to extract from one folder and train from another. They are pretty self explanatory. 
Remove the ones you don't use and copy/edit the ones you make changes to. I have multiple convert scripts at any given time.

Not related, but 2 other tips I'll swear by:

Back your model up. often. Can be as simple as copying and renaming the directory.

Also, when converting, I take a subset of about 300-500 images from the full scene, throw then in a scene_test folder and convert those as a test. Only takes a few minutes and really lets me know how the model training is going
I really appreciate the thorough explanation !
Thanks for the tutorial, i just have one question. does longer duration video need many iteration to be perfected? example i have 10sec and 10min data_dst. if it is 80000 iteration does the quality result from both video will be the same? or the 10min requires more iteration to be perfected?
  •  Previous
  • 1
  • ...
  • 106
  • 107
  • 108(current)
  • 109
  • 110
  • ...
  • 128
  • Next 

Forum Jump:

Users browsing this thread: 5 Guest(s)