diff --git a/doc/faq/faq.md b/doc/faq/faq.md index 5eab320..5a0120d 100644 --- a/doc/faq/faq.md +++ b/doc/faq/faq.md @@ -36,7 +36,8 @@ This is the normal deepfake training process, where src faceset is the celebrity If you are familiar with DeepFaceLab, then this tutorial will help you: -Src faceset is celebrity. Must be diverse enough. +Src faceset is celebrity. Must be diverse enough in yaw, light and shadow conditions. +Src faceset should be xseg'ed and applied. You can apply Generic XSeg to src faceset. Dst faceset is RTM WF faceset from the torrent. @@ -44,7 +45,10 @@ Make a backup before every stage ! > Using SAEHD model. -res:224, WF, ae_dims:256, e_dims:64, d_dims:64, d_mask_dims 22, eyes_mouth_prio:Y, batch 8. Others are default. +res:224, WF, ae_dims:256, e_dims:64, d_dims:64, d_mask_dims 22, eyes_mouth_prio:Y, batch more is better. Others are default. + +Assuming 1kk iters with batch 8. If the batch is higher, iters is possible less. + 1) enable pretrain mode. Train to 1kk 2) disable pretrain mode. Train to 1kk 3) lrd:N uniform_yaw:True, color_transfer:lct, train +500..800k @@ -56,12 +60,20 @@ You can reuse this model to train new src faceset. In this case you should to de > Using AMP model. -res:224, WF, ae_dims:256, inter_dims:1024, e_dims:64, d_dims:64, d_mask_dims:22, morph factor:0.5, batch 8. Others are default. + +Do not mix different ages of people in the src dataset, otherwise the model will approximate age and make the celebrity older/younger depending on the input person. + +res:224, WF, ae_dims:256, inter_dims:1024, e_dims:64, d_dims:64, d_mask_dims:22, morph factor:0.5, batch more is better. Others by default. + +Make a backup before every stage ! + +Assuming 1kk iters with batch 8. If the batch is higher, iters is possible less. 1) lrd:Y, train src-src for 1kk iters -2) delete inter_dst, lrd:N, color_transfer:lct, train +1kk -3) lrd:Y, train +1kk -4) enable gan 0.1 gan_dims:32, train +100..300k iters +2) delete inter_dst, lrd:N, uniform_yaw:True, color_transfer:lct, train +500..1kk +3) color_transfer:none, lrd:Y, train +500..1kk +4) random_warp:False, train +500..1kk +5) enable gan 0.1 gan_dims:32, train +100..300k iters