github/DeepFaceLive
1
0
Fork 0
mirror of https://github.com/iperov/DeepFaceLive synced 2025-08-14 10:47:00 -07:00
Code Issues Projects Releases Packages Wiki Activity Actions

update user_faq training instructions

Browse source
This commit is contained in:
iperov 2022-03-08 13:28:39 +04:00
commit 6b58fc7268
1 changed files with 67 additions and 18 deletions
Download patch file Download diff file Expand all files Collapse all files

85
doc/user_faq/user_faq.md
View file

@ -57,30 +57,58 @@ If you are novice, learn all about DeepFaceLab https://mrdeepfakes.com/forums/th
Gather 5000+ samples of your face with various conditions using webcam which will be used for Live. The conditions are as follows: different lighting, different facial expressions, head direction, eyes direction, being far or closer to the camera, etc. Sort faceset by best to 2000.
> Using SAEHD model.
Here public storage https://disk.yandex.ru/d/7i5XTKIKVg5UUg with facesets and models.
res:224, WF, archi:liae-udt, ae_dims:512, e_dims:64, d_dims:64, d_mask_dims:32, eyes_mouth_prio:Y, blur_out_mask:Y, uniform_yaw:Y, lr_dropout:Y, random_hsv_power:0.1, batch:8. Others by default.
> Using pretrained "RTT model 224.zip" from public storage (see above)
Make a backup before every stage !
1) train +500.000 with RTM WF faceset from the torrent as dst, deleting inter_AB.npy every 100k (save, delete, continue run)
1. place RTM WF Faceset from public storage (see above) to workspace/data_dst/aligned
2) train +500.000
2. place your celeb to workspace/data_src/aligned
3) place your faceset to dst
3. do not change settings. Train +500.000
4) do not delete anything, continue train +500.000
4. replace dst faceset with your faceset in workspace/data_dst/aligned
5) random_warp:OFF, train +500.000
5. continue train +500.000
6) enable gan 0.1 gan_dims:32, train +300.000
6. random_warp:OFF, train +500.000
7. export the model in .dfm format for use in DeepFaceLive. You can also try ordering a deepfake model from someone in Discord or forum.
7. GAN 0.1 power, patch size 28, gan_dims:32. Train until the src loss value has not increased in the last 12 hours.
8. finalize model by disabling masked training for 100-200 (not thousand) iterations.
> Using SAEHD model from scratch.
res:224, WF, archi:liae-udt, ae_dims:512, e_dims:64, d_dims:64, d_mask_dims:32, eyes_mouth_prio:Y, blur_out_mask:Y, uniform_yaw:Y, lr_dropout:Y, batch:8. Others by default.
Make a backup before every stage !
1. place RTM WF Faceset from public storage (see above) to workspace/data_dst/aligned
2. place your celeb to workspace/data_src/aligned
3. train +500.000 deleting inter_AB.npy every 100.000 (save, delete, continue run)
4. train +500.000
5. place your faceset to workspace/data_dst/aligned
6. do not delete anything, continue train +500.000
7. random_warp:OFF, train +500.000
8. GAN 0.1 power, gan_dims:32, Train until the src loss value has not increased in the last 12 hours.
9. finalize model by disabling masked training for 100-200 (not thousand) iterations.
10. export the model in .dfm format for use in DeepFaceLive. You can also try ordering a deepfake model from someone in Discord or forum.
</td></tr>
<tr><td colspan=2 align="center">
## I want to train ready-to-use face model to swap any face to celebrity. What I need to do?
## I want to train ready-to-use face model to swap any face to celebrity, same as public face model. What I need to do?
</td></tr>
<tr><td colspan=2 align="left">
@ -88,29 +116,50 @@ Make a backup before every stage !
If you are familiar with DeepFaceLab, then this tutorial will help you:
Src faceset is celebrity. Must be diverse enough in yaw, light and shadow conditions.
Do not mix different age. The best result is obtained when the face is filmed from a short period of time and does not change the makeup and structure.
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.
> Using pretrained "RTT model 224.zip" from public storage (see above)
> Using SAEHD model.
Make a backup before every stage !
1. place RTM WF Faceset from public storage (see above) to workspace/data_dst/aligned
2. place your celeb to workspace/data_src/aligned
3. place model folder to workspace/model
4. do not change settings, train +500.000 iters
5. random_warp OFF, train +500.000, periodically (every 100.000 iters) disable masked training for 5.000 iters and enable again
6. GAN 0.1 power, patch size 28, gan_dims:32. Train until the src loss value has not increased in the last 12 hours.
7. finalize model by disabling masked training for 100-200 (not thousand) iterations.
> Using SAEHD model from scratch
res:224, WF, archi:liae-udt, ae_dims:512, e_dims:64, d_dims:64, d_mask_dims:32, eyes_mouth_prio:Y, blur_out_mask:Y, uniform_yaw:Y, lr_dropout:Y, batch:8. Others by default.
Make a backup before every stage !
1) train +2.000.000 iters with RTM WF faceset from the torrent as dst, deleting inter_AB.npy every 500k (save, delete, continue run)
1. place RTM WF Faceset from public storage (see above) to workspace/data_dst/aligned
2) random_warp still ON, train +500.000
2. place your celeb to workspace/data_src/aligned
3) if swapped face looks more like dst, delete inter_AB, repeat from stage 2
3. train +2.000.000 iters, deleting inter_AB.npy every 100.000-500.000 iters (save, delete, continue run)
4) random_warp:OFF, train +500.000
4. random_warp still ON, train +500.000
5) enable gan 0.1 gan_dims:32, train +800.000
5. random_warp:OFF, train +500.000
6. GAN 0.1 power, gan_dims:32. Train until the src loss value has not increased in the last 12 hours.
7. finalize model by disabling masked training for 100-200 (not thousand) iterations.
> reusing trained SAEHD RTM model
Models that are trained without random_warp:OFF (before stage 4), can be reused. In this case you have to delete INTER_AB.NPY from the model folder and continue training from stage 2. Increase stage 2 up to 2.000.000 and more iters. You can delete inter_AB.npy every 500.000 iters to increase src-likeness. Trained model before random_warp:OFF also can be reused for new celeb face.
Models that are trained before random_warp:OFF, can be reused. In this case you have to delete INTER_AB.NPY from the model folder and continue training from stage where random_warp:ON. Increase stage up to 2.000.000 and more iters. You can delete inter_AB.npy every 500.000 iters to increase src-likeness. Trained model before random_warp:OFF also can be reused for new celeb face.
</td></tr>
<tr><td colspan=2 align="left">