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.

[DISCUSSION] [SUPPORT] DeepFaceLab 2.0 - Discussion and Support

Status
Not open for further replies.

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
DeepFaceLab 2.0 - Discussion, tips, suggestions - CLOSED

Because no one was using this thread anymore and people prefer to just spam in the guide thread this thread is now closed, if you need to ask about something regarding DFL or the guide, do so in the guide but please use the search option before you post something, chances are your question has been answered.

1. Questions like "how many iterations to achieve quality like some mediocre youtuber deepfaker" are not allowed, the process isn't that simple and such questions are pointless and will be removed and users warned. You can only ask about workflows regarding new options that people are still learning how to correctly implement, new models, etc.
2. No newbie/begginner questions related to basic features, read the guide, then read the FAQs, ALL OF IT, the answers are there.
3. No asking for features to be added to DFL here or anywhere on the forum. In fact don't ask for features to be added to any software here at all, this is not what this forum is about.
4. The guide thread should still be primarly used to ask questions regarding the guide itself, that is too basic explanations, grammar errors, requests to explain new features and optionally DFL bugs reporting. If your question needs more than 1-2 lines to explain the issue and you suspect the answer might be more complicated, create a new thread here in the questions section, if you find a thread that deals in similar issues post in that, don't create a new one just because there is one small difference in what you're dealing with.

DeepFaceLab 2.0 Guide: DFL 2.0 Guide/FAQ

DeepFaceLab 2.0 Github: https://github.com/iperov/DeepFaceLab

Older discussion about DFL 1.0 can be found here: https://mrdeepfakes.com/forums/thread-deepfacelab-explained-and-usage-tutorial

DFL 1.0 is no longer supported, don't use it.

All useful links (guides, sharing and discussion threads): https://mrdeepfakes.com/forums/thread-guide-making-deepfakes-useful-guides-and-threads
 

TMBDF

Moderator | Deepfake Creator | Guide maintainer
Staff member
Moderator
Verified Video Creator
RESERVED FOR FUTURE USE
 

TMBDF

Moderator | Deepfake Creator | Guide maintainer
Staff member
Moderator
Verified Video Creator
aymanalz said:
Can I still do the merging (ie, converting) on an AMD card, after it was trained on google colab's Nvidia?

AFAIK No.

blackyou said:
Nice work. I'm waiting for colab version. Hehe

https://colab.research.google.com/github/chervonij/DFL-Colab/blob/master/DFL_Colab.ipynb

Hanqaqa said:
So happy to try it! Although I still dont know where to download the 2.0 version.

Great job you guys

Don't thank us, thank @iperov unless you want to thank someone for making tutorials and all of that, then sure, you're welcome ;)

Download link is where it always was, on github/google drive, also it's in the 2nd post in this thread.
 

TMBDF

Moderator | Deepfake Creator | Guide maintainer
Staff member
Moderator
Verified Video Creator
jknight said:
tutsmybarreh said:
jknight said:
Getting one OOM after loading models and packs. After that no OOM's. I do however get OOM's and full stop if I turn OFF learn mask. This seems odd as I assume that this would cause less VRAM usage. Anyone else have this issue?


You changed your dims settings, therefore no one will help you, first test your models on default dims and check results, if you feel like the quality isn't as good as you'd want you can try increasing them later.
This requires a lot of testing (to determine impact on quality and how tasking on hardware these changes are).

From my initial tests on 6GB raising 256 to 512 (regular DF) while keeping rest at 64 made me drop batch size by 2 but that was around batch 10-8, if you can run higher batches (or lower) batch required my change by more or less than 2, with DFHD it may be even more of a difference, you also dropped e/d dims and raised d_mask dims too and that makes it even harder to evaluate.

Just run it with default values, also learned mask adds quite a bit to the iteration time and requires lower batch size to run without OOM, test for at least 12 hours, if it doesn't crash after that, then it'll probably run ok. And enable gradient clipping for (insert either fucks or gods) sake unless you don't value your time and like to waste energy by training models that will collapse the second you leave your pc to let it train in the background, it also has an impact on performance and possible batch size but not much, otherwise enable automatic backups.

I understand the lack of support with the modification of dims. I'm just trying to see how far I could push DFL2.0 compared to DFL1.0 at 256res with my card.

I was just curious if anyone has started training with learn_mask True with no OOMS or errors and then switched learned_mask False and having OOMs and full stop when starting training. It just seems like backwards logic to me as I'd expect learn_mask True would require more vram.

And thanks for mentioning clip guard. I guess I forgot to turn it on after mashing the enter key in options.

edit: I am also unable to enable auto backups in options. It just skips it and automatically sets it as no.

On the 1st start yes, but if you then save and start it up again it will let you change which I mentioned in previous posts.
 

kilerb

DF Vagrant
Anyone else having trouble with the new h265 mp4 exports? A few issues I've had are that they won't play in Windows, (VLC player works though.) Maybe a codec download will fix that? It's asking for money in Windows. No biggie... 2nd thing is I re-imported the outputted file into Adobe Premiere and it isn't even close to being synced with the audio (However it does play synced in VLC... So maybe it's another codec issue?) And the 3rd thing is weird. I uploaded my outputted video to my google drive like I always do, so I can grab it on my iphone and look at it. I go to the file and save it (I think these are a bit larger than the h264 files) and it says it's downloading to my iPhone. Never ends up in my photo album. Which I've been doing successfully with all previous version of DFL... Plenty of space on my iPhone. Not sure if it's because it doesn't recognize it as a video or if it's something else on my end, but like I said, never happened before. Thanks!

**Edit** Was just looking again in Premiere. It works fine at the beginning of the video. Audio is synced. Around halfway through it seems to glitch and goes off sync majorly. It's a 3 minute video btw.... And again, it does play in VLC player fine. So not sure what's up there.

**Edit 2** Still playing around on my iPhone. Google drive has processed the video now, and I can play it in Google drive. However, I tried again after processing to download it to my phone. It says "downloading" for a decent amount of time and then stops as usual, but again... Nothing in my photo/video album. So weird.
 

TMBDF

Moderator | Deepfake Creator | Guide maintainer
Staff member
Moderator
Verified Video Creator
kilerb said:
Anyone else having trouble with the new h265 mp4 exports?  A few issues I've had are that they won't play in Windows, (VLC player works though.)  Maybe a codec download will fix that?  It's asking for money in Windows.  No biggie...  2nd thing is I re-imported the outputted file into Adobe Premiere and it isn't even close to being synced with the audio (However it does play synced in VLC...  So maybe it's another codec issue?)  And the 3rd thing is weird.  I uploaded my outputted video to my google drive like I always do, so I can grab it on my iphone and look at it.  I go to the file and save it (I think these are a bit larger than the h264 files) and it says it's downloading to my iPhone.  Never ends up in my photo album.  Which I've been doing successfully with all previous version of DFL...  Plenty of space on my iPhone.  Not sure if it's because it doesn't recognize it as a video or if it's something else on my end, but like I said, never happened before.  Thanks!

**Edit**  Was just looking again in Premiere.  It works fine at the beginning of the video.  Audio is synced.  Around halfway through it seems to glitch and goes off sync majorly.  It's a 3 minute video btw....  And again, it does play in VLC player fine.  So not sure what's up there.

**Edit 2**  Still playing around on my iPhone.  Google drive has processed the video now, and I can play it in Google drive.  However, I tried again after processing to download it to my phone.  It says "downloading" for a decent amount of time and then stops as usual, but again...  Nothing in my photo/video album.   So weird.

If some site is asking for money for "H265" codec then you are either getting scammed or you got some nice malware on your PC already.

I'll test it later once I train my first fake but not all players can play H265 but most should, same goes for some editing software, it just must be supported.

If you're on PC you can try editing VideoEd file in F:\DF\DFL2.0\_internal\DeepFaceLab\mainscripts and change line from
output_kwargs.update ({"c:v": "libx265",
to
output_kwargs.update ({"c:v": "libx264",

I think it should work and then convert the sequence to h264 mp4 file. You can edit it with IDLE. I can't test it myself now.
 

M43r

DF Vagrant
Hey thanks for DFL v2, just wanted to know when the guide would be out ? also can i run this on linux ? (i do have a EGPU) so if can make it work i would let you know !
 

kilerb

DF Vagrant
tutsmybarreh said:
kilerb said:
Anyone else having trouble with the new h265 mp4 exports?  A few issues I've had are that they won't play in Windows, (VLC player works though.)  Maybe a codec download will fix that?  It's asking for money in Windows.  No biggie...  2nd thing is I re-imported the outputted file into Adobe Premiere and it isn't even close to being synced with the audio (However it does play synced in VLC...  So maybe it's another codec issue?)  And the 3rd thing is weird.  I uploaded my outputted video to my google drive like I always do, so I can grab it on my iphone and look at it.  I go to the file and save it (I think these are a bit larger than the h264 files) and it says it's downloading to my iPhone.  Never ends up in my photo album.  Which I've been doing successfully with all previous version of DFL...  Plenty of space on my iPhone.  Not sure if it's because it doesn't recognize it as a video or if it's something else on my end, but like I said, never happened before.  Thanks!

**Edit**  Was just looking again in Premiere.  It works fine at the beginning of the video.  Audio is synced.  Around halfway through it seems to glitch and goes off sync majorly.  It's a 3 minute video btw....  And again, it does play in VLC player fine.  So not sure what's up there.

**Edit 2**  Still playing around on my iPhone.  Google drive has processed the video now, and I can play it in Google drive.  However, I tried again after processing to download it to my phone.  It says "downloading" for a decent amount of time and then stops as usual, but again...  Nothing in my photo/video album.   So weird.

If some site is asking for money for "H265" codec then you are either getting scammed or you got some nice malware on your PC already.

I'll test it later once I train my first fake but not all players can play H265 but most should, same goes for some editing software, it just must be supported.

If you're on PC you can try editing VideoEd file in F:\DF\DFL2.0\_internal\DeepFaceLab\mainscripts and change line from
output_kwargs.update ({"c:v": "libx265",
to
output_kwargs.update ({"c:v": "libx264",

I think it should work and then convert the sequence to h264 mp4 file. You can edit it with IDLE. I can't test it myself now.
I agree that it should play.  But I have 3 computers and all of them need VLC player for 265 files.  I don't think they have viruses.  One of them is really new.  I'll have to investigate more.
 

sharital

DF Admirer
Verified Video Creator
tutsmybarreh said:
frosty3907 said:
I'd taken a timeout to give the program some time to mature... guess it's time to jump back in!

Wait a bit, there will be more updates shortly so don't train to much ;)

Are you saying that models using 2.0 will likely not be compatible with future releases?  Trying to decide to stick with version 1 for a while until 2 matures, or try out 2.  I hate re-training models from scratch, so don't want to have to do it too often. :)
 

TMBDF

Moderator | Deepfake Creator | Guide maintainer
Staff member
Moderator
Verified Video Creator
sharital said:
tutsmybarreh said:
frosty3907 said:
I'd taken a timeout to give the program some time to mature... guess it's time to jump back in!

Wait a bit, there will be more updates shortly so don't train to much ;)

Are you saying that models using 2.0 will likely not be compatible with future releases?  Trying to decide to stick with version 1 for a while until 2 matures, or try out 2.  I hate re-training models from scratch, so don't want to have to do it too often. :)

Models from 1.0 are not compatible with 2.0 and new models trained on 2.0 might be rendered useless by future updates, that's just the nature of in-development software. You can wait or you can start working on it, I doubt there will be any model breaking changes in the near future.


kilerb said:
I agree that it should play.  But I have 3 computers and all of them need VLC player for 265 files.  I don't think they have viruses.  One of them is really new.  I'll have to investigate more.
I wasn't saying VLC is a virus... I was saying you shouldn't be paying for h265, reffering to what you wrote: "Maybe a codec download will fix that?  It's asking for money in Windows.  No biggie..."
 

sharital

DF Admirer
Verified Video Creator
tobe2d said:
I just downloaded DeepFaceLab_NVIDIA_build_01_23_2020. is this  2.0 version?

Yep.




Question for the experts. ;)

In 2.0, in terms of plain-language anticipated results, what is the difference between using DFL and DFLHD if you are using 128 resolution on both?  Typically "HD" refers to a higher resolution, but if you are using a 128x128 pic on both, then it isn't really higher res...... so what does DFLHD do that is "HD" that DFL does not in 2.0?  Thanks in advance for the explanation.

Purpose:  trying to figure out if it is worth it to run DFLHD over DFL even though DFLHD is slower and can run less batches.
 

TMBDF

Moderator | Deepfake Creator | Guide maintainer
Staff member
Moderator
Verified Video Creator
sharital said:
tobe2d said:
I just downloaded DeepFaceLab_NVIDIA_build_01_23_2020. is this  2.0 version?

Yep.




Question for the experts. ;)

In 2.0, in terms of plain-language anticipated results, what is the difference between using DFL and DFLHD if you are using 128 resolution on both?  Typically "HD" refers to a higher resolution, but if you are using a 128x128 pic on both, then it isn't really higher res...... so what does DFLHD do that is "HD" that DFL does not in 2.0?  Thanks in advance for the explanation.

Purpose:  trying to figure out if it is worth it to run DFLHD over DFL even though DFLHD is slower and can run less batches.




There is no "DFLHD", I think you're reffering to different architectures such as DF/LIAE and DFHD/LIAEHD - answer to this you can find in the 3rd post (1st page) in this thread.
If you want to know what's the real difference well, for that you'd have to ask Iperov but it's probably what I wrote, it only reffers to the fact that HD versions are what was SAEHD in the old one and DF is SAE but with features of SAEHD (sort of what @fappenberg modded into his SAE+ mod). It's something to do with additional layers in the neural network.
 

sharital

DF Admirer
Verified Video Creator
Yeah, I meant DFHD vs DF, not DFLHD. Just added an extra L in there.

So basically the answer is "DFHD has mysterious internal stuff that makes it better that only Iperov knows and we don't."

I'll just continue to play around with them. Right now I think I'm forming the opinion that even the new DFHD is too slow on a GTX 1070, so probably will try/use 2.0 DF. (Was previously using SAE+ mod in version 1.)
 

TMBDF

Moderator | Deepfake Creator | Guide maintainer
Staff member
Moderator
Verified Video Creator
sharital said:
Yeah, I meant DFHD vs DF, not DFLHD.  Just added an extra L in there.

So basically the answer is "DFHD has mysterious internal stuff that makes it better that only Iperov knows and we don't."  

I'll just continue to play around with them.  Right now I think I'm forming the opinion that even the new DFHD is too slow on a GTX 1070, so probably will try/use 2.0 DF. (Was previously using SAE+ mod in version 1.)

No, it's not mysterious, you just have to know how neural networks and machine learning works to understand it, I don't know it perfectly either but I've read a bit about it and talked with @fappenberg and really the only difference between DF/LIAE and their HD counterparts (now) is number of layers in the network, rest is the same, before SAE and SAEHD was separate and DF in SAE didn't have all the features and was working in a different way than DF in SAEHD but now they share almost everything except for that thing. As I said we already had this (expect for layers) hacked in SAE+, the reason why it now works better is that iperov further optimised the whole thing so it runs faster.

If you want run HD variants, they should offer more realistic results but DF is still a good option if you don't have the power or want to train your fakes faster. And on that let's end all debates about differences between two, there is no reason to spam it more and I think I explained it as good as I can, the only one who can give you better answer is Iperov but he is busy so I wouldn't bother him with these question.
 

B05283

DF Vagrant
jknight said:
================== Model Summary ==================
==                                               ==
==            Model name: m_SAEHD                ==
==                                               ==
==     Current iteration: 22563                  ==
==                                               ==
==---------------- Model Options ----------------==
==                                               ==
==            autobackup: False                  ==
== write_preview_history: False                  ==
==           target_iter: 0                      ==
==           random_flip: False                  ==
==            resolution: 256                    ==
==             face_type: f                      ==
==     models_opt_on_gpu: False                  ==
==                 archi: dfhd                   ==
==               ae_dims: 512                    ==
==                e_dims: 28                     ==
==                d_dims: 28                     ==
==           d_mask_dims: 28                     ==
==            learn_mask: True                   ==
==            lr_dropout: False                  ==
==           random_warp: True                   ==
==    true_face_training: False                  ==
==      face_style_power: 0.0                    ==
==        bg_style_power: 0.0                    ==
==               ct_mode: none                   ==
==              clipgrad: False                  ==
==              pretrain: False                  ==
==            batch_size: 4                      ==
==                                               ==
==----------------- Running On ------------------==
==                                               ==
==          Device index: 0                      ==
==                  Name: GeForce RTX 2080 SUPER ==
==                  VRAM: 8.00GB                 ==
==                                               ==
===================================================

Getting one OOM after loading models and packs. After that no OOM's. I do however get OOM's and full stop if I turn OFF learn mask. This seems odd as I assume that this would cause less VRAM usage. Anyone else have this issue?

Thanks for sharing a starting point to fiddle around with for RTX 8 GB GPUs. Keen to share your pretrained model?
 

Deeperer

DF Vagrant
Definitely going to give this a try, both CUDA and mutli-CPU.

Request: Could you add a parameter that specifies how many CPU cores will be used? When running the OpenCL / CPU-only version it appears the default number of CPUs invoked by TensorFlow is still 8. I've tried some suggested hacks to config files in the past but they only increased the number of CPUs used for about half of the .bat functions. My CPU-only virtual machine has 24 CPU cores available to it so it would be wonderful to be able to utilize all of them.

Thank you for all your amazing work!
 
Status
Not open for further replies.
Top