Mr DeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
KamroozDownside for having too many source files?
A few hopefully quick questions to gather my head around how training is handled as well as ram used. Is there a downside for using all frames instead of just doing either like 5 or 10fps for extraction from src?

Is there a limit to how many src files one can train on? Would having lets say 6000 images in src images be counter productive?

In regards to ram, if I have more src images, are these all loaded to gpu ram and therefore reducing the amount of settings I can train with? or?

Working on a super project right now and going to train this mofo to perfection. So just trying to wrap my head around all aspects of src images.
Having a larger number of facesets will take longer to train to your desired quality.

It is recommended NOT to do full frames for data_src facesets because most of the frames will be duplicates. Having more images will not reduce the settings you can use because the program groups images into batches when processing.

Most important thing about data_src is to make sure the landmarks are appropriate, the images are clear, and you have the angles required to match data_dst
Further to that- the latest versions do actually seem to get better results with larger source facesets- the only time I honestly think a larger face set is worse is if you're using them from multiple sources/lighting situations. That way you end up with a kind of blurred average of the varied characteristics between the sets.

Forum Jump:

Users browsing this thread: 1 Guest(s)