[SHARING] DFL 2.0 ReadyToTrain model

[INeedMoney]

"Biscuit10" Before you started training on your src's and dst, did you delete inter_b, inter_ab, both? How diverse is you dst?

 

[Biscuit10]

i deleted the Inter_B as mentioned on the front-page.

It was kind of diverse i guess? it was over 1000 images that i have collatedly made aligned_dsts from over several months of learning DFL. About 500 where of the same girl though, but it seems to work really well with any dsts really. Obviously certain facial structures work better from what SRC you have, but i got lucky that the SRC pictures i used (not actual videos broken out) was quite diverse. I have had less fortune with a difference SRC but i think that is because i just have less angles / high quality images to work from.

The only thing i do miss is the TrueFace  thing from the other model-type, that seems to be quite powerful and is the cherry on the top for sure. That seems to bring out the more distinguishing features wherein DST / SRC have similar faces where sometimes its hard to detach from the DST if that makes sense.

I don't know how well that way works with swapping SRCs with the RTM element though.

I could probably also train some of the things for longer too to try get better results, but the loss decrease has stagnated a bit.

I'm going to do some more model learning with the existing 2Million on Florence Pugh and see how that goes. I previously had face extracts that were 2048, but i think it slowed it down too much and i probably don't need it that resolution of face extract for the 256 RTM model on here.

Hope that helps!

 

[tonetantillo]

ReadyToMerge model can be used for:

1) swap the face without the training

2) continue training on the new dst and improve the quality

3) reused to train new fake from scratch. In this case you have to delete INTER_B.npy from the model folder. Also don't forget to set training settings to initial, such as RW(ON), LRD(off), GAN(0.0), etc... Don't enable/disable pretraining !

4) reused to train new RTM model. In this case you have to delete INTER_AB.npy from the model folder.

----------------
The generalization with poor dst faceset will be better than training with default pretrained model.
Example using Tom Cruise model to train Keanu Reeves fake on poor dst faceset from the game:

----------------

Download: <a href='/literotica/magnet:?xt=urn:btih:e8e1d016de49977edd64d04798428bf1057b5e3a&dn=DeepFaceLabReadyToMergeModels&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337%2fannounce'>magnet link</a>

available celebs:

Tom Cruise (224, liae-ud)
Margot Robbie(224, liae-ud)

How to create RTM model from scratch.

Using SAEHD model.
res:224, WF, ae_dims:256, e_dims:64, d_dims:64, d_mask_dims 22, eyes_mouth_prio:Y, batch 8. Others are default.
1) enable pretrain mode. Train to 1kk
2) disable pretrain mode, or use already trained RTM model and delete INTER_AB.npy file . Train to 1kk
3) lrd:N uniform_yaw:True, color_transfer:lct, train +500..800k
4) lrd:Y train +500..800k iters
5) disable random warp +500..800k iters
6) enable gan 0.1 gan_dims:32, train +100..300k iters

1) enable pretrain mode. Train to 1kk 
1000 or 100000?

 

[anirban0104]

ReadyToMerge model can be used for:

1) swap the face without the training

2) continue training on the new dst and improve the quality

3) reused to train new fake from scratch. In this case you have to delete INTER_B.npy from the model folder. Also don't forget to set training settings to initial, such as RW(ON), LRD(off), GAN(0.0), etc... Don't enable/disable pretraining !

4) reused to train new RTM model. In this case you have to delete INTER_AB.npy from the model folder.

----------------
The generalization with poor dst faceset will be better than training with default pretrained model.
Example using Tom Cruise model to train Keanu Reeves fake on poor dst faceset from the game:

----------------

Download: <a href='/literotica/magnet:?xt=urn:btih:e8e1d016de49977edd64d04798428bf1057b5e3a&dn=DeepFaceLabReadyToMergeModels&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337%2fannounce'>magnet link</a>

available celebs:

Tom Cruise (224, liae-ud)
Margot Robbie(224, liae-ud)

How to create RTM model from scratch.

Using SAEHD model.
res:224, WF, ae_dims:256, e_dims:64, d_dims:64, d_mask_dims 22, eyes_mouth_prio:Y, batch 8. Others are default.
1) enable pretrain mode. Train to 1kk
2) disable pretrain mode, or use already trained RTM model and delete INTER_AB.npy file . Train to 1kk
3) lrd:N uniform_yaw:True, color_transfer:lct, train +500..800k
4) lrd:Y train +500..800k iters
5) disable random warp +500..800k iters
6) enable gan 0.1 gan_dims:32, train +100..300k iters

1) enable pretrain mode. Train to 1kk 
1000 or 100000?

1kk = 1 million (1000000)

 

[tonetantillo]

ReadyToMerge model can be used for:

1) swap the face without the training

2) continue training on the new dst and improve the quality

3) reused to train new fake from scratch. In this case you have to delete INTER_B.npy from the model folder. Also don't forget to set training settings to initial, such as RW(ON), LRD(off), GAN(0.0), etc... Don't enable/disable pretraining !

4) reused to train new RTM model. In this case you have to delete INTER_AB.npy from the model folder.

----------------
The generalization with poor dst faceset will be better than training with default pretrained model.
Example using Tom Cruise model to train Keanu Reeves fake on poor dst faceset from the game:

----------------

Download: <a href='/literotica/magnet:?xt=urn:btih:e8e1d016de49977edd64d04798428bf1057b5e3a&dn=DeepFaceLabReadyToMergeModels&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337%2fannounce'>magnet link</a>

available celebs:

Tom Cruise (224, liae-ud)
Margot Robbie(224, liae-ud)

How to create RTM model from scratch.

Using SAEHD model.
res:224, WF, ae_dims:256, e_dims:64, d_dims:64, d_mask_dims 22, eyes_mouth_prio:Y, batch 8. Others are default.
1) enable pretrain mode. Train to 1kk
2) disable pretrain mode, or use already trained RTM model and delete INTER_AB.npy file . Train to 1kk
3) lrd:N uniform_yaw:True, color_transfer:lct, train +500..800k
4) lrd:Y train +500..800k iters
5) disable random warp +500..800k iters
6) enable gan 0.1 gan_dims:32, train +100..300k iters

1) enable pretrain mode. Train to 1kk 
1000 or 100000?

1kk = 1 million (1000000)

Thanks!

 

[iperov]

can you please not to multi quote ?
how to read that shit?

 

[DeliveryBoy]

Mr Iperov thank you so much for this Margot Robbie RTM Model, it's the best model i have ever used and i just freaking love it so much.

 

[iperov]

cool. i am working on better RTM model based on AMP.

 

[Biscuit10]

Is the principle of the AMP model the same as the existing RTM model? Get really good model generalisation against similar to what is done with the pretrain facepack sets, and then do the other big training against your source?

Using yours / DeliveryBoy's RTM models then swapping in own source gets results quite quick, so am curious if it will functionally work the same.

Appreciate it is a new test and answers might not be clear yet of course.

Currently 127k Iterations on SRC-SRC on the pretraining facepack set using the same set-up etc.

 

[chinasilva]

"iperov"
1. your release two models are all Liae models .
2.but when i set the same mask type ,it shows very different results .Margot Robbie can easily dual with glasses and other occlusion .But Tom Cruise.
3. I want to know how to do it ,just like Margot Robbie Model ,I guess you use different Dst training data Can you share your experience?
Thank u very much

 

[slyfox8900]

Am I missing something? The torrent didnt have the DATA SRC facesets for margot or tom. Where can I find them? Thanks

 

[Tratimp]

Am I missing something? The torrent didnt have the DATA SRC facesets for margot or tom. Where can I find them? Thanks

You will not find it in this thread. You are better off trying to find it through the "Celebrity Facesets" section. This thread only has models that allows for instant merging for both Margot and Tom, without the need of data src facesets.

 

[slyfox8900]

Am I missing something? The torrent didnt have the DATA SRC facesets for margot or tom. Where can I find them? Thanks

You will not find it in this thread. You are better off trying to find it through the "Celebrity Facesets" section. This thread only has models that allows for instant merging for both Margot and Tom, without the need of data src facesets.

Oh thank you I didnt realize it works to just go straight to the mergeseahd.dat step. Thanks!

 

[slyfox8900]

"DeliveryBoy" I keep getting out of memory error when using your RTM model with my own SRC Faceset. Im trying to understand where to start with this because even after reading this entire thread I dont quite understand what I need to do to utilize these to make my own RTM Model.

What I understand is I download the RTM model, yours for example (the 2mill iteration one)
I use my own SRC faceset, lets say my own face for example.
I start the train SEAHD, i turn off LRD and turnon RW, and turn off GAN
Batch size to 4
everything else the same

I get instant out of mem error.

Could you clarify the steps for me to get my own face to be ready for RTM starting from your shared RTM model?

Im using RTX 3080 FYI, so I dont understand. Also for some reason the training terminal says my card only has 8.25gbs of memory. I swore in the past it said 10gbs but idk whats changed.

 

[Biscuit10]

what resolution are you attempting with the SRC / DST? The model is 256, so if the src / dst are larger its making it do extra work for no gain, and larger SRC takes longer iterations.

I would start with the base 2million iterations and start with LRD / GAN off to see if you can do batch 4 without it to start with, if it loads, great - quit out then turn the settings back on. Leave RW off in this instance too if you're just testing it. Primary focus is the reason of OOM.

Also open up the task mananger to check your GPU levels

 

[rjSampaio]

HI everyone, im being dabling on DFL for a 2 weeks and found out this is probably the best forum to ask questions as this tool as so many options and lingo that im overwhelm, my +15years on programing do not help at all

I found out yesterday about this post and im a confuse about the applications.

I have my own source and destination, would this train model help me get better results, as its better than to pretrain myself? I im running source/destination of my friends and families, unfortunately Tom/Margot are not part of it.  

3) reused to train new fake from scratch. In this case you have to delete INTER_B.npy from the model folder. Also don't forget to set training settings to initial, such as RW(ON), LRD(off), GAN(0.0), etc... Don't enable/disable pretraining !

4) reused to train new RTM model. In this case you have to delete INTER_AB.npy from the model folder.

In the points above, number 3 when saying "new fake" is what? a new destination? in what senario should i delete the _B only?

In number 4, that's what i want if i have my own source and destination? 
and do i need to also delete _B 

Sorry for the noob questions, as i struggle to find a video tutorial regarding this.

 

[knightkat]

"DeliveryBoy" I keep getting out of memory error when using your RTM model with my own SRC Faceset. Im trying to understand where to start with this because even after reading this entire thread I dont quite understand what I need to do to utilize these to make my own RTM Model.

What I understand is I download the RTM model, yours for example (the 2mill iteration one)
I use my own SRC faceset, lets say my own face for example.
I start the train SEAHD, i turn off LRD and turnon RW, and turn off GAN
Batch size to 4
everything else the same

I get instant out of mem error.

Could you clarify the steps for me to get my own face to be ready for RTM starting from your shared RTM model?

Im using RTX 3080 FYI, so I dont understand. Also for some reason the training terminal says my card only has 8.25gbs of memory. I swore in the past it said 10gbs but idk whats changed.

I think you need to change your page swap size.
Control Pannel -> System -> Advance System Settings -> Advanced -> Settings -> Advanced -> Change -> Custom Size
I upped my maximum size to 10000 and it seems to work.

 

[pluviophile]

nvm

 

[reneholy]

did anyone try training with many dst faces using DF model?

 

[moekali]

could someone help me with this error pls ?