Training method tips

[turnip26Participant]

Hi, I’ve been using Deepfacelab for some time now and have produced some decent results. I just wondered what your method is for training. Do you have a set process? I have a 1080Ti so use 256 or 320 resolution and I always use a pre-trained model (a million iterations) but which options should I turn on at which iteration count to refine my model and improve the output quality?

It would be useful to know which settings make the most improvements, particularly those below.

== uniform_yaw: False ==
== lr_dropout: n ==
== random_warp: False ==
== gan_power: 0.0 ==
== true_face_power: 0.0 ==
== face_style_power: 0.0 ==
== bg_style_power: 0.0 ==
== clipgrad: True ==
== random_flip: True ==
== eyes_mouth_prio: False ==
== blur_out_mask: True ==
== adabelief: True ==
== random_hsv_power: 0.0 ==
== random_src_flip: False ==
== random_dst_flip: True ==
== gan_patch_size: 40 ==
== gan_dims: 16

[Author]

Posts