DeepFaceLive Tutorial

Table of Contents

0:25 – Channel Intro
0:38 – Going over prerequisites/requirements
2:32 – Mentioning the RTT Model
3:10 – Discussing various GPU’s which would be ideal for training an RTT DF Live model
4:10 – DF Lab now works with AMD cards but at the time I made this video they had not yet fixed it – I may address how to fix it in a later video
5:27 – Files you’ll need (see Description for Download Links)
6:25 – Discussing the RTM Faceset
10:59 – Extracting DeepfaceLab, going over what to keep and what to discard | Faceset and model files
18:03 – Extracting images from DATA_SRC.mp4
20:44 – Deleting SRC images which do not contain your SRC character
23:10 – Extracting faces from source imagery (Whole Face)
25:40 – How to quickly (more or less) delete bad alignments or extracted faces which are not good quality, or are not your SRC character
31:45 – Checking out the extracted faces after deleting all unwanted images
34:25 – Applying Generic Xseg mask & training
37:29 – Generic Xseg was insufficient; Manually masking required
38:03 – Manually Xseg masking Jim/Ernest
41:43 – Results of training after manual Xseg’ing was added to Generically trained mask
43:03 – Applying Xseg training to SRC
43:45 – Archiving our SRC faces into a “faceset.pak” archive file for faster loading times
47:40 – Beginning training of our SAEHD model
51:00 – Color transfer mode bugs out and throws python errors
51:29 – Training successfully launches without color transfer mode (enabled later in the video successfully)
51:10 – Explaining the immediately evident benefits of training with the RTT model
55:00 – Saving the model and training on my 3090 Ti instead
56:49 – The next morning after ~9 hours of training (124k iterations)
58:25 – Changing settings to enable Learning Rate Dropout (LRD)
1:00:27 – 264k iterations, shutting off Random Warp (RW)
1:03:56 – 405k iterations, enabling GAN (microphone misbehaving, sorry)
1:08:51 – Discussing GAN visual “gridlike” artifacting which is normal
1:10:27 – Showing GAN “mottled” appearance (ie visual artifacting) – ~505k iterations
1:15:19 – Enabled “SOT” Color Transfer mode off camera
1:15:33 – Showing the improvements made by SOT mode, ~575k iterations
1:21:00 – 633k iterations (Final Product, but one more day of training would’ve probably helped)
1:24:37 – Extracting DeepFace Live software, going over what files are included
1:26:50 – Exporting the SAEHD DFM file of Jim Varney
1:28:55 – Copying the DFM file over to the DeepFace Live model folder
1:29:50 – Running DeepfaceLive
1:32:00 – Camera bugs out because it’s still enabled in OBS – Removing the camera from OBS solves the conflict
1:33:10 – Enabling Jim Varney in DeepFace Live after picking settings
1:36:14 – Final comments