MrDeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
mrrobotSOLVED Some frames being skipped when converting
#1
Hi all,
  I'm fairly new to DFL.  I'm creating a pretty simple fake where I'm swapping my face to Henry Cavill's from Superman.  I'm using SAEHD


Code:
================ Model Summary =================
==                                            ==
==            Model name: SAEHD               ==
==                                            ==
==     Current iteration: 327074              ==
==                                            ==
==-------------- Model Options ---------------==
==                                            ==
==            autobackup: True                ==
== write_preview_history: True                ==
==           random_flip: True                ==
==            resolution: 128                 ==
==             face_type: f                   ==
==            learn_mask: True                ==
==        optimizer_mode: 1                   ==
==                 archi: df                  ==
==               ae_dims: 256                 ==
==            ed_ch_dims: 21                  ==
==            lr_dropout: True                ==
==           random_warp: True                ==
==    true_face_training: True                ==
==      face_style_power: 0.2                 ==
==        bg_style_power: 0.2                 ==
==               ct_mode: rct                 ==
==              clipgrad: False               ==
==            batch_size: 8                   ==
==                                            ==
==---------------- Running On ----------------==
==                                            ==
==          Device index: 0                   ==
==                  Name: GeForce GTX 1080 Ti ==
==                  VRAM: 11.00GB             ==
==                                            ==
================================================

  So far I have 371,228 iterations with loss values of .1639/.1163

  My question is this - when I use convert SAEHD.bat some of the frames are being skipped.  I've removed all the crap faces from my dataset before training my model and data_dst has every frame from the clip in it.  Some of the frames that are being skipped are almost identical to frames that are being converted.  Is there a step that I might be missing that will fix this issue for me?  Here are the steps I've taken:

1. Extract images from data_src
2. Extract images from data_dst
3. Extract all faces from data_src
4. Extract all faces from data_dst

5. data_src sort by x
6. data_src check results
7. remove bad frames
8. repeat 5-7 until happy with dataset
9. data_dst sort by x
10. data_dst check results
11. remove bad frames
12. repeat 9-11 until happy with results
13. train SAE HD using above settings
14. convert SAE HD
15. covert to mp4

tiya!
#2
(01-14-2020, 07:22 PM)mrrobot Wrote: You are not allowed to view links. Register or Login to view.Hi all,
  I'm fairly new to DFL.  I'm creating a pretty simple fake where I'm swapping my face to Henry Cavill's from Superman.  I'm using SAEHD


Code:
================ Model Summary =================
==                                            ==
==            Model name: SAEHD               ==
==                                            ==
==     Current iteration: 327074              ==
==                                            ==
==-------------- Model Options ---------------==
==                                            ==
==            autobackup: True                ==
== write_preview_history: True                ==
==           random_flip: True                ==
==            resolution: 128                 ==
==             face_type: f                   ==
==            learn_mask: True                ==
==        optimizer_mode: 1                   ==
==                 archi: df                  ==
==               ae_dims: 256                 ==
==            ed_ch_dims: 21                  ==
==            lr_dropout: True                ==
==           random_warp: True                ==
==    true_face_training: True                ==
==      face_style_power: 0.2                 ==
==        bg_style_power: 0.2                 ==
==               ct_mode: rct                 ==
==              clipgrad: False               ==
==            batch_size: 8                   ==
==                                            ==
==---------------- Running On ----------------==
==                                            ==
==          Device index: 0                   ==
==                  Name: GeForce GTX 1080 Ti ==
==                  VRAM: 11.00GB             ==
==                                            ==
================================================

  So far I have 371,228 iterations with loss values of .1639/.1163

  My question is this - when I use convert SAEHD.bat some of the frames are being skipped.  I've removed all the crap faces from my dataset before training my model and data_dst has every frame from the clip in it.  Some of the frames that are being skipped are almost identical to frames that are being converted.  Is there a step that I might be missing that will fix this issue for me?  Here are the steps I've taken:

1. Extract images from data_src
2. Extract images from data_dst
3. Extract all faces from data_src
4. Extract all faces from data_dst

5. data_src sort by x
6. data_src check results
7. remove bad frames
8. repeat 5-7 until happy with dataset
9. data_dst sort by x
10. data_dst check results
11. remove bad frames
12. repeat 9-11 until happy with results
13. train SAE HD using above settings
14. convert SAE HD
15. covert to mp4

tiya!

Well you need to check data_dst/aligned using 5.1) data_dst check results debug to see if there are frames without aligned faces if you see frames with incorrect landmarks or no ladmarks at all you must delete that frame and then align them manually with 5) data_dst extract faces MANUAL RE-EXTRACT DELETED RESULTS DEBUG. Otherwise if there is no face detected for that frame it will output original one with any faces swapped onto it.
If I helped you in any way or you enjoy my deepfakes, please consider a small donation.
Bitcoin: 1C3dq9zF2DhXKeu969EYmP9UTvHobKKNKF
Want to request a paid deepfake or have any questions reagarding the forums or deepfake creation using DeepFaceLab? Write me a message.
TMB-DF on the main website - You are not allowed to view links. Register or Login to view.

Forum Jump:

Users browsing this thread: 1 Guest(s)