Mr DeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
FlameFirst fake, a little feedback
#1
First training on intro scene. Only clear faces.
Trained 40-50k with :
 
===== Model summary =====
== Model name: SAE
==
== Current iteration: 0
==
== Model options:
== |== batch_size : 4
== |== sort_by_yaw : False
== |== random_flip : False
== |== resolution : 128
== |== face_type : f
== |== learn_mask : True
== |== optimizer_mode : 1
== |== archi : df
== |== ae_dims : 512
== |== e_ch_dims : 42
== |== d_ch_dims : 21
== |== multiscale_decoder : True
== |== ca_weights : False
== |== pixel_loss : False
== |== face_style_power : 10.0
== |== bg_style_power : 10.0
== |== apply_random_ct : True
== Running on:
== |== [0 : GeForce GTX 1080]
 
Next pushed batch_size to 8 with optimize mode on 2 :
 
== Model options:
== |== batch_size : 8
== |== sort_by_yaw : False
== |== random_flip : False
== |== resolution : 128
== |== face_type : f
== |== learn_mask : True
== |== optimizer_mode : 2
== |== archi : df
== |== ae_dims : 512
== |== e_ch_dims : 42
== |== d_ch_dims : 21
== |== remove_gray_border : False
== |== multiscale_decoder : True
== |== pixel_loss : True
== |== face_style_power : 0.1
== |== bg_style_power : 4.0
== Running on:
== |== [0 : GeForce GTX 1080]
 
Quite nice result. Waited to reach 80k total.
 
Next extracted 2nd scene with more movement and trained with these settings :
 
== Model options:
== |== batch_size : 4
== |== sort_by_yaw : False
== |== random_flip : False
== |== resolution : 128
== |== face_type : f
== |== learn_mask : True
== |== optimizer_mode : 1
== |== archi : df
== |== ae_dims : 512
== |== e_ch_dims : 42
== |== d_ch_dims : 21
== |== multiscale_decoder : True
== |== ca_weights : False
== |== pixel_loss : False
== |== face_style_power : 0.1
== |== bg_style_power : 0.4
== |== apply_random_ct : True
== Running on:
== |== [0 : GeForce GTX 1080]
 
Face looks less merged ? But still decent result. Trained one night about 55k or 60k iterations.
Then pushed batch size to 12 with optimizer mode 2 and following settings :
 
== Model options:
== |== batch_size : 12
== |== sort_by_yaw : False
== |== random_flip : False
== |== resolution : 128
== |== face_type : f
== |== learn_mask : True
== |== optimizer_mode : 2
== |== archi : df
== |== ae_dims : 512
== |== e_ch_dims : 42
== |== d_ch_dims : 21
== |== multiscale_decoder : True
== |== ca_weights : False
== |== pixel_loss : False
== |== face_style_power : 0.1
== |== bg_style_power : 4.0
== |== apply_random_ct : True
== Running on:
== |== [0 : GeForce GTX 1080]
 
Face now looks a little more like dst than src. But it's ok we still recognize src.
Currently at 152k and pretty good result (on preview).
 
Conclusion : Looks like beginning training with face_style_power 10 and bg_style_power 10 is the best choice.
Since decreasing it for 1st pass on second scene produced a less merged face. Also batch_size 12 looks a lot better to me for only a 300ms increase in iterations compared to batch 8. (maybe placebo but ATM seems worth). Also decreased both loss values a much more than batch_size 8.
 Anyway good result at this stage. Probably going to wait 160k and try converting the firsts two scenes to then move to training on more "intense" scenes … Smile
 
Questions : What is exactly the link between batch_size and quality ? What is the link between face_style_power/bg_style_power values and merging of src to dst ? Should I train straight up to batch 12 instead of starting at 4, even if iterations are x2 longer ? Does this workflow looks ok to more experienced deepfakers ?
 
Makes a lot of questions but even with carefully reading documentation I can't really answer them… I can include few previews if needed.

Thanks in advance to deepfakes that will reply to this thread.

fLmmm
#2
Higher batch will give a better result because it runs the proces against more faces
#3
(05-13-2019, 09:52 PM)GhostTears Wrote: You are not allowed to view links. Register or Login to view.Higher batch will give a better result because it runs the proces against more faces

Thanks for your answer ! 
And what about bg_style_power and face_style_power ? I undestand that they have something to do with colors but I don't really know how to modify these values throughout the training.
I began conversion and saw that the mask is clearly noticeable on most of the merged pictures. I believe it has something to do with these settings but which one ? (I have backups so I can revert)
It wasn't happening before 2nd pass on second scene with bg_style_power 4.
#4
think about style power as of morphing the faces. Higher face style power = more it will morph to look like data_dst and the same with background (everything outside the mask).

Style power also fixes skin tone issues because you're making it look more like data_dst. But now there's also RCT during transfer so it reduces the need.

I generally add more power if the data_src face shape is different than data_dst... But yes, adding more power will make it look less like data_src.
#5
(05-14-2019, 02:41 AM)dpfks Wrote: You are not allowed to view links. Register or Login to view.think about style power as of morphing the faces. Higher face style power = more it will morph to look like data_dst and the same with background (everything outside the mask).

Style power also fixes skin tone issues because you're making it look more like data_dst. But now there's also RCT during transfer so it reduces the need.

I generally add more power if the data_src face shape is different than data_dst... But yes, adding more power will make it look less like data_src.

Okay ! I had it wrong x) 
BTW I uploaded my first converted scene : You are not allowed to view links. Register or Login to view.. 
I wont upload second scene because I am not hapy with it. The face is changing colours every 3-4 seconds which makes it "blinking" a little. Don't know what I did wrong. Maybe RCT during conversion. Don't know. 
But I'm pretty happy with the one linked above. I am open to any sugestion to make it better. 
Conversion settings all on default. Seamless mode + fan-DST mask. No erode no blur or anything.

Forum Jump:

Users browsing this thread: 1 Guest(s)