MrDeepFakes Forums

Some content may not be available to Guests. Consider registering an account to enjoy unrestricted access to guides, support and tools

  • We are looking for community members who are intested in helping out. See our HELP WANTED post.

Any traıned model for SAEHD

Status
Not open for further replies.

Mr.BaBe

DF Enthusiast
Verified Video Creator
Hello

on older version of DFL I was using trained models, and those models are very successful. Any chance to see SAEHD trainlen models here soon?

Thanks in advance
 

kkdlux

DF Admirer
I have a pretrained model saehd -170 resolution 75k iterations and now i´m doing a 160 resolution DF, currently at 55k.

Custom random pictures added to the pretrain folder with better quality and some from 3d models.

lpojEEQ.png
 

kkdlux

DF Admirer
a50ZW4j.png


https://mega.nz/#F!E8ADQIoL!eojy0JxKBnN6KB2ZiCoRNw

=============== Model Summary ===============
== ==
== Model name: SAEHD ==
== ==
== Current iteration: 73002 ==
== ==
==------------- Model Options -------------==
== ==
== sort_by_yaw: True ==
== random_flip: True ==
== resolution: 160 ==
== face_type: f ==
== learn_mask: True ==
== optimizer_mode: 3 ==
== archi: df ==
== ae_dims: 512 ==
== ed_ch_dims: 21 ==
== random_warp: False ==
== true_face_training: False ==
== face_style_power: 0.0 ==
== bg_style_power: 0.0 ==
== ct_mode: False ==
== clipgrad: False ==
== batch_size: 12 ==
== ==
==-------------- Running On ---------------==
== ==
== Device index: 0 ==
== Name: GeForce GTX 1080 Ti ==
== VRAM: 11.00GB ==
== ==
=============================================
 

kkdlux

DF Admirer
=============== Model Summary ===============
== ==
== Model name: SAEHD ==
== ==
== Current iteration: 82484 ==
== ==
==------------- Model Options -------------==
== ==
== sort_by_yaw: True ==
== random_flip: True ==
== resolution: 176 ==
== face_type: f ==
== learn_mask: True ==
== optimizer_mode: 2 ==
== archi: liae ==
== ae_dims: 256 ==
== ed_ch_dims: 21 ==
== random_warp: False ==
== true_face_training: False ==
== face_style_power: 0.0 ==
== bg_style_power: 0.0 ==
== ct_mode: False ==
== clipgrad: False ==
== batch_size: 8 ==
== ==
==-------------- Running On ---------------==
== ==
== Device index: 0 ==
== Name: GeForce GTX 1080 Ti ==
== VRAM: 11.00GB ==
== ==
=============================================

For me it works better than the DF one.

https://mega.nz/#F!gghR1SZK!59aCTRfHJPvJ370hHTIa3g
 

Mr.BaBe

DF Enthusiast
Verified Video Creator
Thanks a lot for the model file. Can you share some pictures to show the difference between DF and Liae, I will try this one personally, but never used liae before. I am sure some comprehension pictures will help many newbies like me.
 

kkdlux

DF Admirer
With DF the face doesn´t transfer as fast, it keeps the second face almost the same, the LIAE one is faster and i pretrained the model in a different way.

eF2mzD5.png



And this is the first one on DF

tTHDrQ2.png
 

vogmqdlfc

DF Pleb
kkdlux said:
=============== Model Summary ===============
== ==
== Model name: SAEHD ==
== ==
== Current iteration: 82484 ==
== ==
==------------- Model Options -------------==
== ==
== sort_by_yaw: True ==
== random_flip: True ==
== resolution: 176 ==
== face_type: f ==
== learn_mask: True ==
== optimizer_mode: 2 ==
== archi: liae ==
== ae_dims: 256 ==
== ed_ch_dims: 21 ==
== random_warp: False ==
== true_face_training: False ==
== face_style_power: 0.0 ==
== bg_style_power: 0.0 ==
== ct_mode: False ==
== clipgrad: False ==
== batch_size: 8 ==
== ==
==-------------- Running On ---------------==
== ==
== Device index: 0 ==
== Name: GeForce GTX 1080 Ti ==
== VRAM: 11.00GB ==
== ==
=============================================

For me it works better than the DF one.

https://mega.nz/#F!gghR1SZK!59aCTRfHJPvJ370hHTIa3g

WHAT WAS IT PER?..........thx for the model
 

kkdlux

DF Admirer
stron01 said:
Thanks a lot for the model file. Can you share some pictures to show the difference between DF and Liae, I will try this one personally, but never used liae before. I am sure some comprehension pictures will help many newbies like me.

vogmqdlfc said:
kkdlux said:
=============== Model Summary ===============
==                                         ==
==         Model name: SAEHD               ==
==                                         ==
==  Current iteration: 82484               ==
==                                         ==
==------------- Model Options -------------==
==                                         ==
==        sort_by_yaw: True                ==
==        random_flip: True                ==
==         resolution: 176                 ==
==          face_type: f                   ==
==         learn_mask: True                ==
==     optimizer_mode: 2                   ==
==              archi: liae                ==
==            ae_dims: 256                 ==
==         ed_ch_dims: 21                  ==
==        random_warp: False               ==
== true_face_training: False               ==
==   face_style_power: 0.0                 ==
==     bg_style_power: 0.0                 ==
==            ct_mode: False               ==
==           clipgrad: False               ==
==         batch_size: 8                   ==
==                                         ==
==-------------- Running On ---------------==
==                                         ==
==       Device index: 0                   ==
==               Name: GeForce GTX 1080 Ti ==
==               VRAM: 11.00GB             ==
==                                         ==
=============================================

For me it works better than the DF one.

https://mega.nz/#F!gghR1SZK!59aCTRfHJPvJ370hHTIa3g

WHAT WAS IT PER?..........thx for the model

IT around 1400 ms
 

vogmqdlfc

DF Pleb
kkdlux said:
stron01 said:
Thanks a lot for the model file. Can you share some pictures to show the difference between DF and Liae, I will try this one personally, but never used liae before. I am sure some comprehension pictures will help many newbies like me.

vogmqdlfc said:
kkdlux said:
=============== Model Summary ===============
==                                         ==
==         Model name: SAEHD               ==
==                                         ==
==  Current iteration: 82484               ==
==                                         ==
==------------- Model Options -------------==
==                                         ==
==        sort_by_yaw: True                ==
==        random_flip: True                ==
==         resolution: 176                 ==
==          face_type: f                   ==
==         learn_mask: True                ==
==     optimizer_mode: 2                   ==
==              archi: liae                ==
==            ae_dims: 256                 ==
==         ed_ch_dims: 21                  ==
==        random_warp: False               ==
== true_face_training: False               ==
==   face_style_power: 0.0                 ==
==     bg_style_power: 0.0                 ==
==            ct_mode: False               ==
==           clipgrad: False               ==
==         batch_size: 8                   ==
==                                         ==
==-------------- Running On ---------------==
==                                         ==
==       Device index: 0                   ==
==               Name: GeForce GTX 1080 Ti ==
==               VRAM: 11.00GB             ==
==                                         ==
=============================================

For me it works better than the DF one.

https://mega.nz/#F!gghR1SZK!59aCTRfHJPvJ370hHTIa3g

WHAT WAS IT PER?..........thx for the model

IT around 1400 ms

with K80 i am getting 6000 ms.....using google collab
 

vogmqdlfc

DF Pleb
vogmqdlfc said:
kkdlux said:
stron01 said:
Thanks a lot for the model file. Can you share some pictures to show the difference between DF and Liae, I will try this one personally, but never used liae before. I am sure some comprehension pictures will help many newbies like me.

vogmqdlfc said:
kkdlux said:
=============== Model Summary ===============
==                                         ==
==         Model name: SAEHD               ==
==                                         ==
==  Current iteration: 82484               ==
==                                         ==
==------------- Model Options -------------==
==                                         ==
==        sort_by_yaw: True                ==
==        random_flip: True                ==
==         resolution: 176                 ==
==          face_type: f                   ==
==         learn_mask: True                ==
==     optimizer_mode: 2                   ==
==              archi: liae                ==
==            ae_dims: 256                 ==
==         ed_ch_dims: 21                  ==
==        random_warp: False               ==
== true_face_training: False               ==
==   face_style_power: 0.0                 ==
==     bg_style_power: 0.0                 ==
==            ct_mode: False               ==
==           clipgrad: False               ==
==         batch_size: 8                   ==
==                                         ==
==-------------- Running On ---------------==
==                                         ==
==       Device index: 0                   ==
==               Name: GeForce GTX 1080 Ti ==
==               VRAM: 11.00GB             ==
==                                         ==
=============================================

For me it works better than the DF one.

https://mega.nz/#F!gghR1SZK!59aCTRfHJPvJ370hHTIa3g

WHAT WAS IT PER?..........thx for the model

IT around 1400 ms

with K80 i am getting 6000 ms.....using google collab

within 5000 its....dst looks good.......so far.............why didnt u try res 128.....then try trueface later on?
 

Mr.BaBe

DF Enthusiast
Verified Video Creator
You can not change the resolution, but you can turn on trueface anytime. With CPU i5 and rtx 2070 cideo card (8gb vgram), I can run trainer with this resolution, with learning, true face and wrap on, as long as powers set to 0.0.
I use my src for different dst. Every time I run trueface and wrap for 2-3k ite, and then run at least 1-2k only learning mask on and the results are very good. I think regardless of what, on this new version, learning mask should be turned on all the time.
Also I am using upscaled src pictures with topaz gigapixel (1024/1024 pictures). it also give an edge to final video.
with higher resolution like this one and training a bit trueface and wrap shows wrinkles on the face more visible.
 
Status
Not open for further replies.
Top