[SHARING] DFL 2.0 ReadyToTrain model

[schweinebier]

320 Res, LIAE-UD Pre-Train Model, using Groggy's Female Faceset, will restart training on this to 1M in another week or two, had to pause. Working towards this thread for the RTM Model for a specific faceset. First pre-train model from scratch so if there's anything wrong with the dims my bad
resolution: 320
face_type: wf
archi: liae-ud
ae_dims: 256
e_dims: 64
d_dims: 64
d_mask_dims: 22
Current Iterations: 573,200

I tried to use this model with my src but it starts from iteration 1, what am I doing wrong? Copied all files in my model folder and disabled the pretrain mode...
I tried iperovs model beforehand - this worked fine - strange it is

 

[INeedMoney]

Why are the landmarks on the full jawline in the dst faceset, out of curiosity?
Would this DST set still be ideal for making nsfw rtm's?

 

[texapo9999]

320 Res, LIAE-UD Pre-Train Model, using Groggy's Female Faceset, will restart training on this to 1M in another week or two, had to pause. Working towards this thread for the RTM Model for a specific faceset. First pre-train model from scratch so if there's anything wrong with the dims my bad
resolution: 320
face_type: wf
archi: liae-ud
ae_dims: 256
e_dims: 64
d_dims: 64
d_mask_dims: 22
Current Iterations: 573,200

I tried to use this model with my src but it starts from iteration 1, what am I doing wrong? Copied all files in my model folder and disabled the pretrain mode...
I tried iperovs model beforehand - this worked fine - strange it is

Iperovs was trained with Tom Cruise already, if you check his steps for making your own RTM model, this is just a higher resolution than the recommended 224. However, another user here shared one with higher dims trained to 1M iterations that I was able to use instead

 

[INeedMoney]

Question: Say you train a SRC with this RTM DST. Then you delete inter_b and train that model with a more specific DST (i.e. pornstars only, or even just a 1-pornstar DST). Will the model lose it's adaptability?

I'm working on this now. I'll make backups when I switch DSTs, and do some comparison merges. I'm curious what the experts already know.

 

[alucardvii]

Question: Say you train a SRC with this RTM DST. Then you delete inter_b and train that model with a more specific DST (i.e. pornstars only, or even just a 1-pornstar DST). Will the model lose it's adaptability?

Uh, I don't think you delete inter_b when switching DSTs. I think you would only delete it if you switch SRCs (I'm not 100% sure).

 

[INeedMoney]

"alucardvii"

INTER_B holds info about DST faces and it's a part of LIAE architecture. As said you should delete that file since it will re-train DST faces that way the model can adapt to new DST.

I don't know. It sounds like we should delete it when we change DST. Maybe I shouldn't have deleted it since I only switched to another multi-person dst and not a final DST?

 

[TMBDF]

Question: Say you train a SRC with this RTM DST. Then you delete inter_b and train that model with a more specific DST (i.e. pornstars only, or even just a 1-pornstar DST). Will the model lose it's adaptability?

Uh, I don't think you delete inter_b when switching DSTs. I think you would only delete it if you switch SRCs (I'm not 100% sure).

The point of RTM model is that you change just DST, never SRC, that always remains the same, thus you delete inter file when changing DST.

Just read the instructions on the first page, they are simple and straightforward, it's all explained, none of this disucssion is even needed here (no offence, but this is a sharing thread, for discussions there are other threads).

 

[DeliveryBoy]

Hey guys i have pretrained model that is 256 resolution, done already 2 million iterations with the following settings

================== Model Summary ===================
==                                                                       ==
==            Model name: RTM Liae-UD-256_SAEHD   ==
==                                                ==
==     Current iteration: 2000000                  ==
==                                                ==
==---------------- Model Options -----------------==
==                                                ==
==            resolution: 256                     ==
==             face_type: wf                      ==
==     models_opt_on_gpu: True                    ==
==                 archi: liae-ud                 ==
==               ae_dims: 256                     ==
==                e_dims: 64                      ==
==                d_dims: 64                      ==
==           d_mask_dims: 22                      ==
==       masked_training: True                    ==
==       eyes_mouth_prio: True                    ==
==           uniform_yaw: True                    ==
==             adabelief: True                    ==
==            lr_dropout: n                       ==
==           random_warp: False                   ==
==       true_face_power: 0.0                     ==
==      face_style_power: 0.0                     ==
==        bg_style_power: 0.0                     ==
==               ct_mode: lct                     ==
==              clipgrad: False                   ==
==              pretrain: True                    ==
==       autobackup_hour: 0                       ==
== write_preview_history: False                   ==
==           target_iter: 2000000                 ==
==       random_src_flip: False                   ==
==       random_dst_flip: False                   ==
==            batch_size: 8                       ==
==             gan_power: 0.0                     ==
==        gan_patch_size: 32                      ==
==              gan_dims: 16                      ==
==                                                ==
==------------------ Running On ------------------==
==                                                ==
==          Device index: 0                       ==
==                  Name: NVIDIA GeForce RTX 2080 ==
==                  VRAM: 6.54GB                  ==
==                                                ==
====================================================

What do i have to do now to convert this into Ready To Merge model?

 

[kaemen]

I will def try this out THank YOU!

 

[TMBDF]

Hey guys i have pretrained model that is 256 resolution, done already 2 million iterations with the following settings

================== Model Summary ===================
==                                                                       ==
==            Model name: RTM Liae-UD-256_SAEHD   ==
==                                                ==
==     Current iteration: 2000000                  ==
==                                                ==
==---------------- Model Options -----------------==
==                                                ==
==            resolution: 256                     ==
==             face_type: wf                      ==
==     models_opt_on_gpu: True                    ==
==                 archi: liae-ud                 ==
==               ae_dims: 256                     ==
==                e_dims: 64                      ==
==                d_dims: 64                      ==
==           d_mask_dims: 22                      ==
==       masked_training: True                    ==
==       eyes_mouth_prio: True                    ==
==           uniform_yaw: True                    ==
==             adabelief: True                    ==
==            lr_dropout: n                       ==
==           random_warp: False                   ==
==       true_face_power: 0.0                     ==
==      face_style_power: 0.0                     ==
==        bg_style_power: 0.0                     ==
==               ct_mode: lct                     ==
==              clipgrad: False                   ==
==              pretrain: True                    ==
==       autobackup_hour: 0                       ==
== write_preview_history: False                   ==
==           target_iter: 2000000                 ==
==       random_src_flip: False                   ==
==       random_dst_flip: False                   ==
==            batch_size: 8                       ==
==             gan_power: 0.0                     ==
==        gan_patch_size: 32                      ==
==              gan_dims: 16                      ==
==                                                ==
==------------------ Running On ------------------==
==                                                ==
==          Device index: 0                       ==
==                  Name: NVIDIA GeForce RTX 2080 ==
==                  VRAM: 6.54GB                  ==
==                                                ==
====================================================

What do i have to do now to convert this into Ready To Merge model?

Once you pretrained your model you typically train it.

Don't think about it as some kind conversion, at the end of the day it's just a trick to get the model to work on various faces and never as good as proper training could be on one specific dst.

Simply train it on provided random dst and your src of choice until you get results that you like, then you can go straight into merging on various dsts or delete inter file and train some more on your new dst (don't enable RW when training on new DST once you delete inter).

 

[DeliveryBoy]

Hey guys i have pretrained model that is 256 resolution, done already 2 million iterations with the following settings

================== Model Summary ===================
==                                                                       ==
==            Model name: RTM Liae-UD-256_SAEHD   ==
==                                                ==
==     Current iteration: 2000000                  ==
==                                                ==
==---------------- Model Options -----------------==
==                                                ==
==            resolution: 256                     ==
==             face_type: wf                      ==
==     models_opt_on_gpu: True                    ==
==                 archi: liae-ud                 ==
==               ae_dims: 256                     ==
==                e_dims: 64                      ==
==                d_dims: 64                      ==
==           d_mask_dims: 22                      ==
==       masked_training: True                    ==
==       eyes_mouth_prio: True                    ==
==           uniform_yaw: True                    ==
==             adabelief: True                    ==
==            lr_dropout: n                       ==
==           random_warp: False                   ==
==       true_face_power: 0.0                     ==
==      face_style_power: 0.0                     ==
==        bg_style_power: 0.0                     ==
==               ct_mode: lct                     ==
==              clipgrad: False                   ==
==              pretrain: True                    ==
==       autobackup_hour: 0                       ==
== write_preview_history: False                   ==
==           target_iter: 2000000                 ==
==       random_src_flip: False                   ==
==       random_dst_flip: False                   ==
==            batch_size: 8                       ==
==             gan_power: 0.0                     ==
==        gan_patch_size: 32                      ==
==              gan_dims: 16                      ==
==                                                ==
==------------------ Running On ------------------==
==                                                ==
==          Device index: 0                       ==
==                  Name: NVIDIA GeForce RTX 2080 ==
==                  VRAM: 6.54GB                  ==
==                                                ==
====================================================

What do i have to do now to convert this into Ready To Merge model?

Once you pretrained your model you typically train it.

Don't think about it as some kind conversion, at the end of the day it's just a trick to get the model to work on various faces and never as good as proper training could be on one specific dst.

Simply train it on provided random dst and your src of choice until you get results that you like, then you can go straight into merging on various dsts or delete inter file and train some more on your new dst (don't enable RW when training on new DST once you delete inter).

Thank you for this clear and precise instruction.

Now i see where i kept going wrong, i kept enabling RW on new DST after deleting the inter file.

 

[DeliveryBoy]

Simply train it on provided random dst

The provided random dst? Is that the celeb facepack i used for pretraining?

The file random dst facepak that was provided in the torrent file by the original OP (thread starter) is only 512 resolution i prefer to train on 768 resolution facepak.

 

[TMBDF]

Simply train it on provided random dst

The provided random dst? Is that the celeb facepack i used for pretraining?

The file random dst facepak that was provided in the torrent file by the original OP (thread starter) is only  512 resolution i prefer to train on 768 resolution facepak.

You can use any dataset as long as it's random faces and it covers wide range of facial expressions, angles and lighting conditions, variety is the key.

 

[DeliveryBoy]

Pretraining Model & Ready to Merge Model



WF
Pretraining Model

LIAE-UD - res: 256, dims: 256/64/64/22, 2.000.000 iterations, Whole Face, Facesets 1024x1024  (NSFW - FFHQ - PreTrain)



WF
ReadyTOMerge Model
LIAE-UD - res: 256, dims: 256/64/64/22, 2.000.000 iterations, Whole Face, XSeg Masked, Female 1024x1024 SRC, RANDOM Male and Female 1024x1024 DST   (NSFW - FFHQ - ReadToMerge Trained)

DOWNLOAD:

 

[o12925624]

WF

LIAE-UD - res: 256, dims: 256/64/64/22, 1.301.050 iterations, Whole Face, XSeg Masked, Female Only 1024x1024 SRC, RANDOM 1024x1024 DST   (SFW - FFHQ - ReadToMerge Trained)

DOWNLOAD:

I've tried to ready to merge model you posted, for which thanks.  I put the model files in my folder but after a training for an hour I still don't see the SRC showing up in the combined src/dst preview (its 100% dest).  Does it just take a lot longer or am I doing something wrong?

What I did was:
- Put your files in the model directory with my own src/dst
- Delete RTM Liae-UD-256_SAEHD_inter_B
- Run the model, RW(enabled), LRD(disabled), GAN (0)

Also, could someone please seed the original torrent? I've been trying to get it over the past days but no one seems to be seeding.

 

[Biscuit10]

o12925624 did you make sure you put "Pretain" to N? i downloaded the model to test it and noticed it was on by default.

___

I've tried the method from the front page,  this method is great, absolutely all the thanks ever.

However I'm getting OOM however with LRD & GAN

I had separate models before which had 254 Resolution and i could with below details:

If i go LRD drop to 6, then GAN to 4 - but i can't run the model from first page higher than batch 4 with LRD / GAN turned off (GAN & LRD with Batch 2 also OOMs)

Any thoughts?

p.s. ignore the iteration count, i know it was from an initial Pre-training but the screenshot is still relevant in terms of the model parameters.

 

[o12925624]

o12925624 did you make sure you put "Pretain" to N? i downloaded the model to test it and noticed it was on by default.

___

I downloaded the model posted by because I still haven't been able to get the torrent (no seeders) to train my own models.
This model I'm testing with is already configured as pretain = n

Could someone maybe post the data_dst from the torrent somewhere where I can get it?
Thanks

 

[DeliveryBoy]

o12925624 did you make sure you put "Pretain" to N? i downloaded the model to test it and noticed it was on by default.

___

I downloaded the model posted by because I still haven't been able to get the torrent (no seeders) to train my own models.
This model I'm testing with is already configured as pretain = n

Could someone maybe post the data_dst from the torrent somewhere where I can get it?
Thanks

Data_Dst is a video you want to use for the outcome of the deep fake
Data_Src is the face of someone you want to use for the deep fake in the Data_dst video
Ready To Merth 256 Model is a trained model


Find a video you want to deep fake (Data_Dst), find a face you want to use in Deepfaking (Data_Src), and then you use this ready to merge model to deepfake your work.

I will upload the Data_Dst, Data_src, and a new ready to merge 256 model (1.500,000 interations) for you tomorrow night when i am back in town.

Note: you do not need my Data_Dst or my Data_src i have used, all you need to do is make your own data_dst and data_src, then you use this readytomerge model to make your own deepfake.

Both the data_dst and data_src picture resolution i used were 1024 x 1024 in both the pretraining and the real training.

---

o12925624 did you make sure you put "Pretain" to N? i downloaded the model to test it and noticed it was on by default.

___

I downloaded the model posted by because I still haven't been able to get the torrent (no seeders) to train my own models.
This model I'm testing with is already configured as pretain = n

Could someone maybe post the data_dst from the torrent somewhere where I can get it?
Thanks

Here is the data_dst i used (1024x1024) random faces high resolution males + females ((XSeg Generic) data_dst whole_face mask - applied)
Download Link:

Here is the data_src i used (1024x1024) female ((XSeg Generic) data_src whole_face mask - applied)
Download Link:

Here is the normal pretraining (2,000,000 iterations), 256 Liae-ud, WF, masked_training=True,  eyes_mouth_priority=True, uniform_yaw=True, adabelief=True, High Resolution 1024x1024
if you want to continue with this pretraining, you may.
Download Link:


Here is the latest ReadyToMerge training, done extra training to 1,500,000 iterations.
Download Link:

 

[boobsanlubes]

o12925624 did you make sure you put "Pretain" to N? i downloaded the model to test it and noticed it was on by default.

___

I downloaded the model posted by because I still haven't been able to get the torrent (no seeders) to train my own models.
This model I'm testing with is already configured as pretain = n

Could someone maybe post the data_dst from the torrent somewhere where I can get it?
Thanks

Data_Dst is a video you want to use for the outcome of the deep fake
Data_Src is the face of someone you want to use for the deep fake in the Data_dst video
Ready To Merth 256 Model is a trained model


Find a video you want to deep fake (Data_Dst), find a face you want to use in Deepfaking (Data_Src), and then you use this ready to merge model to deepfake your work.

I will upload the Data_Dst, Data_src, and a new ready to merge 256 model (1.500,000 interations) for you tomorrow night when i am back in town.

Note: you do not need my Data_Dst or my Data_src i have used, all you need to do is make your own data_dst and data_src, then you use this readytomerge model to make your own deepfake.

Both the data_dst and data_src picture resolution i used were 1024 x 1024 in both the pretraining and the real training.

---

o12925624 did you make sure you put "Pretain" to N? i downloaded the model to test it and noticed it was on by default.

___

I downloaded the model posted by because I still haven't been able to get the torrent (no seeders) to train my own models.
This model I'm testing with is already configured as pretain = n

Could someone maybe post the data_dst from the torrent somewhere where I can get it?
Thanks

Here is the data_dst i used (1024x1024) random faces high resolution males + females ((XSeg Generic) data_dst whole_face mask - applied)
Download Link:

Here is the data_src i used (1024x1024) female ((XSeg Generic) data_src whole_face mask - applied)
Download Link:

Here is the normal pretraining (2,000,000 iterations), 256 Liae-ud, WF, masked_training=True,  eyes_mouth_priority=True, uniform_yaw=True, adabelief=True, High Resolution 1024x1024
if you want to continue with this pretraining, you may.
Download Link:


Here is the latest ReadyToMerge training, done extra training to 1,500,000 iterations.
Download Link:

---

Thank you for sharing this. This is going to speed things up during training yes?

 

[DeliveryBoy]

It will speed up your training by a lot lol.