DeepFaceLab Tutorial – Advanced Training Methods

Table of Contents

00:00 – Intro, general info, System requirements, etc.
30:14 – Generic XSEG + training with the 13m iteration XSEG model files
39:54 – Checking out the sorted Anne Hathaway faceset, packing the faceset
44:05 – Creating a new 320 res model
50:59 – Overwriting the encoder/decoder with their RTT equivalents
52:36 – Scrambled dst training images = it worked
55:05 – 600 iterations and already looks good!
55:41 – 11k iterations
57:33 – 243k iterations, enabling RW+LRD
1:00:22 – 435k iterations, shutting off RW (Going LRD “only”)
1:07:49 – 617k, enabling GAN
1:14:25 – 720k, Demonstrating what GAN artifacting/grid patterns look like
1:18:34 – 850k, going from Batch 8 to Batch 32 (because I felt like it, this is not required/necessary)
1:26:42 – 900k, sufficiently trained to export a DFM file and test in DF Live
1:28:43 – DFM successfully exported, file then moved to DF Live folder for accessibility by the DF Live application
1:29:42 – Testing the DFM model in DF Live (not bad, lower teeth need more source examples to define better)
1:34:09 – How to recycle/re-use the model files from Anne Hathaway to create a new source character (Star Trek TNG “Data” in this case)
1:38:08 – 15 minutes of training and Data already looks impressive
1:40:49 – How to insta-merge your RTM trained model into any DST clip (Terminator 3 in this case)
1:42:06 – Showing how well Anne looks on Kristana Loken in Terminator 3 (SAEHD Merge Window), final comments/closing