Mr DeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
BlueroseDeepFaceLab - training issues (blurry)
#1
Greetings from an absolute beginner. I have watched couple of videos on YouTube about creating deep-fakes using DeepFaceLab. It seems pretty straight forward however my process seems to start dwindling when I am training.

 I cant seem to get any clear images while training as everything is blurry. I have over 400 images in my src folder which seem to look ok in the workspace\data_src\aligned directory. All faces are clear (not blurry at all).

I have fiddled with process this using the train H64.bat file. I have even let it run for 30 minutes to train without any success. Both tutorials i watched say something about training until a certain value is reached but it seems my CMD window inst outputting any values for 'loss':

[Image: cmd.jpg]

As you can see I am running a GTX 970 and a quad-core, so my machine seems capable of handling this. Using DeepFaceLabCUDA9.2SSE

I read on the NVIDA site that gtx9780 are supportedfor CUDA, but the DeepFaceLab would not run properly with it installed.  I revereted back to using TensorFlow (default) just to get things going but wondering if there is a better way for me to get good results. 

also.. when I use 7) convert H64.bat i get the error in console "copying without faces" which I am not sure how to rectify.

Any suggestions or help is much appreciated., Thank you for your time.
#2
Train for 3 days, read my guide. Train for over 100k+ iterations. In your screenshot it's only 760.

Copying without faces usually occurs for frames without a face detected. If you always see that then you're doing something wrong.
~ Fake it till you make it ~
You are not allowed to view links. Register or Login to view.
#3
(08-16-2019, 01:39 AM)dpfks Wrote: You are not allowed to view links. Register or Login to view.Train for 3 days, read my guide. Train for over 100k+ iterations. In your screenshot it's only 760.

Copying without faces usually occurs for frames without a face detected. If you always see that then you're doing something wrong.

Thanks for taking the time to respond. I am not sure if you are being serious about training for 3 days. Are you being serious ? i havent read anywhere else that training needs to take this long. XD

As for the 760 iterations, do I increase this value by simply letting the training run longer? Is this why the image is looking blurry?

As for copying without faces goes, yes something is wrong, but how do I correct this? The dst video always has a a few sections where the face is partially obscured by a shoulder and antother section of the clip where the head is partly out of the frame. Could this be why I am seeing this error?

Thanks again!
#4
I train for 5 days. Again, read my guide....

I train to 180,000 iterations.

Like a said, if the scene does not have a face, it will state that message. If you are getting it for all frames then yes something is wrong.
~ Fake it till you make it ~
You are not allowed to view links. Register or Login to view.
#5
(08-16-2019, 04:57 AM)dpfks Wrote: You are not allowed to view links. Register or Login to view.I train for 5 days. Again, read my guide....

I train to 180,000 iterations.

Like a said, if the scene does not have a face, it will state that message. If you are getting it for all frames then yes something is wrong.

ok wow thank you. I will totally follow your guide. Is there any chance you have a link for it?
#6
You are not allowed to view links. Register or Login to view.
~ Fake it till you make it ~
You are not allowed to view links. Register or Login to view.

Forum Jump:

Users browsing this thread: 1 Guest(s)