[SHARING] DFL 2.0 ReadyToTrain model

[Biscuit10]

Using the 1,500,000 model and own DST / SRC, is this bit in the notepad file:

"175k RW -> 60k RW LRD -> 35k no-RW -> 20k no-RW LRD"

On top of the 1,500,000 base for the model?

Cheers!

 

[boobsanlubes]

It will speed up your training by a lot lol.

You uploaded a pretraining (2M iter) and a RTM (1.5M iter)...which one should I use for my custom deepfakes? I'm still confused between the difference of pretraining and RTM.

 

[DeliveryBoy]

It will speed up your training by a lot lol.

You uploaded a pretraining (2M iter) and a RTM (1.5M iter)...which one should I use for my custom deepfakes? I'm still confused between the difference of pretraining and RTM.

The 2 million iteration pretrained model is for when you want to start your own training from scratch.

The RTM 1.5M Iteration is for when you want a much faster result with a lot less iterations.

---

Using the 1,500,000 model and own DST / SRC, is this bit in the notepad file:

"175k RW -> 60k RW LRD -> 35k no-RW -> 20k no-RW LRD"

On top of the 1,500,000 base for the model?

Cheers!

I made my own readtomerge model according to the tutorial on how to make one by the original poster.

If you use the ReadyToMerge model, this is the result you will get really fast:

 

[boobsanlubes]

It will speed up your training by a lot lol.

You uploaded a pretraining (2M iter) and a RTM (1.5M iter)...which one should I use for my custom deepfakes? I'm still confused between the difference of pretraining and RTM.

The 2 million iteration pretrained model is for when you want to start your own training from scratch.

The RTM 1.5M Iteration is for when you want a much faster result with a lot less iterations.

---

Using the 1,500,000 model and own DST / SRC, is this bit in the notepad file:

"175k RW -> 60k RW LRD -> 35k no-RW -> 20k no-RW LRD"

On top of the 1,500,000 base for the model?

Cheers!

I made my own readtomerge model according to the tutorial on how to make one by the original poster.

If you use the ReadyToMerge model, this is the result you will get really fast:

---

I understand now. Thanks!

Seeing that you also trained for just an hour and a half and you already achieved good results...this is amazing.

 

[boobsanlubes]

"DeliveryBoy"

I tried to redownload the file: but it would seems that you have deleted it?

 

[DeliveryBoy]

Sorry my bad i forgot to update the link

---

Using the 1,500,000 model and own DST / SRC, is this bit in the notepad file:

"175k RW -> 60k RW LRD -> 35k no-RW -> 20k no-RW LRD"

On top of the 1,500,000 base for the model?

Cheers!

Dont forget to add GAN after you disabled RW, this will make the skin teeth and eyes very sharp and in focus. You only need to do around 40K iteration for GAN (gan=0.1) Also enable history preview and choose the picture you want for previewing so you could use this 1 picture as a reference, this helps GAN work faster.

To view the history of the GAN, view the first jpg file and then the last jpg file, you will see gan in action

this should be your work flow: "175k RW -> 60k RW LRD -> 35k no-RW -> 20k no-RW LRD -> 40K GAN"

How to view GAN progression: I enabled GAN at 1666680 iteration (this will be first file to view, next there is a file called _last.jpg, you view the _last.jpg file just like in the picture example i have uploaded below). Dont view the rest of the other files ... just view the first file and then the last file so  you could see the huge improvement gan has made on your work.

 

[boobsanlubes]

Sorry my bad i forgot to update the link

---

Using the 1,500,000 model and own DST / SRC, is this bit in the notepad file:

"175k RW -> 60k RW LRD -> 35k no-RW -> 20k no-RW LRD"

On top of the 1,500,000 base for the model?

Cheers!

Dont forget to add GAN after you disabled RW, this will make the skin teeth and eyes very sharp and in focus. You only need to do around 40K iteration for GAN (gan=0.1) Also enable history preview and choose the picture you want for previewing so you could use this 1 picture as a reference, this helps GAN work faster.

To view the history of the GAN, view the first jpg file and then the last jpg file, you will see gan in action

this should be your work flow: "175k RW -> 60k RW LRD -> 35k no-RW -> 20k no-RW LRD -> GAN til at least 40K iterations"

How to view GAN progression: I enabled GAN at 1666680 iteration (this will be first file to view, next there is a file called _last.jpg, you view the _last.jpg file just like in the picture example i have uploaded below). Dont view the rest of the other files ... just view the first file and then the last file so  you could see the huge improvement gan has made on your work.

---

You're a godsend! I tried to download your old mega.nz file with over 9gb overnight when I woke up it says quota exceeded for my I.P and will need to wait for 8 hours. xD

 

[DeliveryBoy]

Sorry my bad i forgot to update the link

---

Using the 1,500,000 model and own DST / SRC, is this bit in the notepad file:

"175k RW -> 60k RW LRD -> 35k no-RW -> 20k no-RW LRD"

On top of the 1,500,000 base for the model?

Cheers!

Dont forget to add GAN after you disabled RW, this will make the skin teeth and eyes very sharp and in focus. You only need to do around 40K iteration for GAN (gan=0.1) Also enable history preview and choose the picture you want for previewing so you could use this 1 picture as a reference, this helps GAN work faster.

To view the history of the GAN, view the first jpg file and then the last jpg file, you will see gan in action

this should be your work flow: "175k RW -> 60k RW LRD -> 35k no-RW -> 20k no-RW LRD -> GAN til at least 40K iterations"

How to view GAN progression: I enabled GAN at 1666680 iteration (this will be first file to view, next there is a file called _last.jpg, you view the _last.jpg file just like in the picture example i have uploaded below). Dont view the rest of the other files ... just view the first file and then the last file so  you could see the huge improvement gan has made on your work.

---

You're a godsend! I tried to download your old mega.nz file with over 9gb overnight when I woke up it says quota exceeded for my I.P and will need to wait for 8 hours. xD

---

You're welcome, you could try waiting, or change your ip address lol.

 

[tonetantillo]

Sorry my bad i forgot to update the link

---

Using the 1,500,000 model and own DST / SRC, is this bit in the notepad file:

"175k RW -> 60k RW LRD -> 35k no-RW -> 20k no-RW LRD"

On top of the 1,500,000 base for the model?

Cheers!

Dont forget to add GAN after you disabled RW, this will make the skin teeth and eyes very sharp and in focus. You only need to do around 40K iteration for GAN (gan=0.1) Also enable history preview and choose the picture you want for previewing so you could use this 1 picture as a reference, this helps GAN work faster.

To view the history of the GAN, view the first jpg file and then the last jpg file, you will see gan in action

this should be your work flow: "175k RW -> 60k RW LRD -> 35k no-RW -> 20k no-RW LRD -> GAN til at least 40K iterations"

How to view GAN progression: I enabled GAN at 1666680 iteration (this will be first file to view, next there is a file called _last.jpg, you view the _last.jpg file just like in the picture example i have uploaded below). Dont view the rest of the other files ... just view the first file and then the last file so  you could see the huge improvement gan has made on your work.

---

You're a godsend! I tried to download your old mega.nz file with over 9gb overnight when I woke up it says quota exceeded for my I.P and will need to wait for 8 hours. xD

---

You're welcome, you could try waiting, or change your ip address lol.

---

DeliveryBoy
What faceset did you use for dst. You mentioned "use dst faceset from the torrent" which one is that
Do you use my own faceset for src? Does it have to be of the same person?

 

[Biscuit10]

Perfect - thank you for your reply.

Last question - do you keep YAW / Eye Prio on throughout when putting a new DST / SRC?

And presumably when i have a SRC that has had the above iterations complete and i just want to change out the DST, i imagine there are some short-cuts that can be taken as the new SRC will be trained on a lot?

Am giving this a test today

 

[DeliveryBoy]

Sorry my bad i forgot to update the link

---

Using the 1,500,000 model and own DST / SRC, is this bit in the notepad file:

"175k RW -> 60k RW LRD -> 35k no-RW -> 20k no-RW LRD"

On top of the 1,500,000 base for the model?

Cheers!

Dont forget to add GAN after you disabled RW, this will make the skin teeth and eyes very sharp and in focus. You only need to do around 40K iteration for GAN (gan=0.1) Also enable history preview and choose the picture you want for previewing so you could use this 1 picture as a reference, this helps GAN work faster.

To view the history of the GAN, view the first jpg file and then the last jpg file, you will see gan in action

this should be your work flow: "175k RW -> 60k RW LRD -> 35k no-RW -> 20k no-RW LRD -> GAN til at least 40K iterations"

How to view GAN progression: I enabled GAN at 1666680 iteration (this will be first file to view, next there is a file called _last.jpg, you view the _last.jpg file just like in the picture example i have uploaded below). Dont view the rest of the other files ... just view the first file and then the last file so  you could see the huge improvement gan has made on your work.

---

You're a godsend! I tried to download your old mega.nz file with over 9gb overnight when I woke up it says quota exceeded for my I.P and will need to wait for 8 hours. xD

---

You're welcome, you could try waiting, or change your ip address lol.

---

DeliveryBoy
What faceset did you use for dst. You mentioned "use dst faceset from the torrent" which one is that
Do you use my own faceset for src? Does it have to be of the same person?

---

i used the one that came with the torrent but i made the resolution 1024 x 1024, i also used my own female src faceset (SFW + NSFW) 1024 x 1024.

You can use your own src and dst

I will update my RTM 256 model with 2,000,000 iterations in a few hours. This version has 800,000 iterations with GAN = 0.1 on. (Pretrained with 2,000,000 iteratioins and RTM trained with 2,000,000 iterations) should be a good model for you guys to play with.

When Nvidia drops the prices of the RTX3000 series, i will buy the 3090 card with 24 gig vram, or better yet the 4090 will be out in 12 months, i might as well wait for the 4090. Then i can share more models with you guys.

I want to attempt doing 10,000,000 iterations for pretraining model then use it to make 10,000,000 iterations for RTM model.

 

[DeliveryBoy]

Pretraining Model & Ready to Merge Model (256 resolution WF Liau-ud)

Updated: 2 files in 1 (Pretrained 2,000,000 iterations + ReadToMerge 2,000,000 iterations.

 

[tonetantillo]

Thanks! DeliveryBoy

 

[luxlethor]

"DeliveryBoy"
I've been trying to download this for the past 6 hours, tried different browsers and whatnot and I just keep getting "temporary error, retrying" -error from mega.
Any change to put this somewhere else too?

EDIT: Working now. Thank you.

 

[boobsanlubes]

"DeliveryBoy"
I've been trying to download this for the past 6 hours, tried different browsers and whatnot and I just keep getting "temporary error, retrying" -error from mega.
Any change to put this somewhere else too?

Same error here.

EDIT:

Working fine and already finished the download.

 

[angel_uriel]

Getting "Temporary error, retrying" also

EDIT:
Working Now

 

[Biscuit10]

Just want to say, massive shout-out to you for the Model Deliveryboy (and ofcourse Iperov in the first instance).

I've taken the 2Million RTM model, taken my various source files to make a model each per source and using a wide variety of data_dst files that i have been practicing and using over the last few months while learning this.

Ran it overnight against 3 sources, one on PC, 2 on Colab and are each around 40k iterations (after the 2mil) and results are looking great. Each are around 0.1500 loss and are still creeping down comfortably (haven't applied GAN yet). PC can't do LRD unfortunately as i OOM, but even with LRD results are looking great, just think LRD would got me there a bit quicker, but works well batch 4 without it.

A good thing too, each source i'm testing with are just regular photos with decent variety (not perfect), but around 200-250 each. With such a small sample size vs the initial 5000+, it's turning out astonishingly well.

DST's being photos works great, but it is also working quite well with short videos despite the small source samples (key is still variety for sure), but backed up by the 2million pre-training & 2 million actual training makes your own source bloody great.

I dont think this thread is linked in the guide, but i believe it should be.

Fantasic work, and all the thanks for sharing.

 

[iperov]

updated main post. Added Margot Robbie model.

 

[DeliveryBoy]

updated main post. Added Margot Robbie model.

thank you Sir

---

updated main post. Added Margot Robbie model.

Could you give us link to your Margo Robbie Data_src please?

 

[anirban0104]

I think we are supposed to take the src from Head and whole_face facesets torrent. But there's only 3k pics of margot