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.

Are these specs good enough to run Deepfakes?

Hey Everyone!

I am looking to purchase a new computer and figured I would get one that is capatible with fakeapp or deepfakes so that I can give this a try.  Does anyone know what I can expect to pay for a Nvidia GPU that is 3.0 or higher? I have a keyboard, mouse, monitor, and power cords.  All I would need is the tower, or laptop with a operation system. 

I found this today while looking, and wanted to double check if these specs are good enough. If somenoe could help me out, I would appreciate it. Looking forward to your guys' feedback. 

GPU: NVidia GTX Geoforce 1050 8GB
Operating System: Windows 10
Graphics Processing Type: Dedicated
CPU: Intel Core i5 3rd Gen. 3.20 GHz

Would this set-up work?
 

Pocketspeed

DF Admirer
Verified Video Creator
If you can afford to upgrade the GPU, I would.  In order to use DFL, you'll need a CUDA compute capability of 3.7.  The Faceswap git will require 3.5.  VRAM is the most important, followed by a fairly recent CPU, and then plenty of hard drive space (some deepfakes I have in progress are 60+ Gb). The new Villain model of the Faceswap git will need 11 Gb VRAM, at least. You can use it with 8 Gb of VRAM, but you will be limited in your settings, and your batch size will be 8 or less. But the Villain model is the best damn thing I've seen so far.

But if you just want to do some basic deepfakes, just make sure your GPU's CUDA compute capability is at least 3.7 and has 8 Gb of VRAM, and you have a multi-threaded CPU that supports AVX and AVX2. And I would suggest available hard drive space of 300 to 500 Gb.
 
Pocketspeed said:
If you can afford to upgrade the GPU, I would.  In order to use DFL, you'll need a CUDA compute capability of 3.7.  The Faceswap git will require 3.5.  VRAM is the most important, followed by a fairly recent CPU, and then plenty of hard drive space (some deepfakes I have in progress are 60+ Gb). The new Villain model of the Faceswap git will need 11 Gb VRAM, at least. You can use it with 8 Gb of VRAM, but you will be limited in your settings, and your batch size will be 8 or less. But the Villain model is the best damn thing I've seen so far.

But if you just want to do some basic deepfakes, just make sure your GPU's CUDA compute capability is at least 3.7 and has 8 Gb of VRAM, and you have a multi-threaded CPU that supports AVX and AVX2. And I would suggest available hard drive space of 300 to 500 Gb.

Thanks for the quick reply! I appreciate the help! So will the above specifications work?
 

Pocketspeed

DF Admirer
Verified Video Creator
They will for the Faceswap git.  I don't use DFL so I'm not sure, but my understanding is that you need a CUDA compute capability of 3.7, and that you'll probably have to use the H64 model with that card.  The GTX 1050 has a CUDA compute of 6.1, so you should be good.
 

jreg72

DF Vagrant
The 1050 has a compute capability of 6.1, so it would technically work. A better card and more VRAM would perform better of course and let you use things like the villain model Pocketspeed mentioned and whatever will come up in the future.
 

iperov

DF Enthusiast
Developer
Pocketspeed said:
If you can afford to upgrade the GPU, I would.  In order to use DFL, you'll need a CUDA compute capability of 3.7.  The Faceswap git will require 3.5.  VRAM is the most important, followed by a fairly recent CPU, and then plenty of hard drive space (some deepfakes I have in progress are 60+ Gb). The new Villain model of the Faceswap git will need 11 Gb VRAM, at least. You can use it with 8 Gb of VRAM, but you will be limited in your settings, and your batch size will be 8 or less. But the Villain model is the best damn thing I've seen so far.

But if you just want to do some basic deepfakes, just make sure your GPU's CUDA compute capability is at least 3.7 and has 8 Gb of VRAM, and you have a multi-threaded CPU that supports AVX and AVX2. And I would suggest available hard drive space of 300 to 500 Gb.

why is Villain model best ? did you compare it with DFL ?

@homerj975  DFL recommended VRAM is 6GB for all models include SAE.
this is trained on 6GB : 
 

Pocketspeed

DF Admirer
Verified Video Creator
@iperov I didn't mean any offense.  DFL gives very good results.  It's just personal preference.  I didn't say it was the best, I said it was the "best I've seen so far." That's just opinion.

Example: 
 

Cerciusx

DF Pleb
What exactly is this villian model? I have just been training against my previous pretrained models with good results.
 

iperov

DF Enthusiast
Developer
I adapted it to SAE .
You can update DFL github folder and test new 'vg' archi.


also SAE now support 64-256 res
 

emsbabs

DF Vagrant
iperov said:
I adapted it to SAE .
You can update DFL github folder and test new 'vg' archi.


also SAE now support 64-256 res



What is the easiest way to update DFL? Download the new one from Github and delete the last one?
I just download the prebuild windows app. Is that the right way to do it or should get something else?
 

emsbabs

DF Vagrant
iperov said:
replace DFL folder in _internal

Sorry that I'm bad at understanding. But on your Github I im not sure what to download, so I'll get the updated version.?
Is it the prebuild windows app or something from Github?

By the way, very good job in DFL. I really admire your work and would love to support a little to your futere RTX 2080 Ti as soon as the paycheck ticks in :)
 

Pocketspeed

DF Admirer
Verified Video Creator
Absolutely no doubt that @iperov is a Grade A dev, and by bringing the villain model into DFL, @iperov has definitively made DFL a DeepFake source to be reckoned with. Not to mention that AMD cards can now be used with DFL, which in my opinion, actually gives DFL a huge lead over the Faceswap git.

At this point, it's purely a matter of personal preference which tool to use. DFL is definitely easier for the new deepfaker, and users are not limited to just nVidia cards.

Just giving respect where respect is due. Very impressive stuff.


@CerciusX I'm generally sticking with the original model. But I do intend to experiment with the Villain model in the future. The Villain model requires a lot of VRAM. You can use it with 8Gb if you are selective in settings, which means I will not be able to use all options. Villain is part of the new model refactor implemented in the most recent commit of the Faceswap master branch repo. The results are pretty damned impressive. High quality, high def, and is able to swap very dissimilar faces. GTX 1080ti is recommended, but 1070ti is doable with a batch size of 8 or less. The only issue with the Faceswap model refactor is that now I have to start my models from the beginning, instead of training over previous models like you do.
 

iperov

DF Enthusiast
Developer
villain now part of SAE, which is very flexible. You can change dimensions and train VG even on 2GB card :D
 

Cerciusx

DF Pleb
I'm having trouble with the setup python script in the faceswap git not seeing my cuDNN install on Linux. tried a bunch of things but haven't got it working. I'll ave to stick to standard models for now I guess. Not a DeepFaceLab user because of all the finickiy stuff I'd have to work out to get it to work on linux. You;d think the dev would cater to linux users but eh. whatev. If these new models become standard I'll probably bow out of the community because I can't compete with a p2000 4GB and no working deepfacelab install for linux.
 

andy_ger

DF Vagrant
emsbabs said:
iperov said:
replace DFL folder in _internal

Sorry that I'm bad at understanding. But on your Github I im not sure what to download, so I'll get the updated version.?
Is it the prebuild windows app or something from Github?

By the way, very good job in DFL. I really admire your work and would love to support a little to your futere RTX 2080 Ti as soon as the paycheck ticks in :)

Hi emsbabs,

you have to install the prebuild windows app once.

After that download the github repository



and then extract the content in the zip to the \_internal\bin\DeepFaceLab directory in your prebuild windows app install dir

 

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
Cerciusx said:
I'm having trouble with the setup python script in the faceswap git not seeing my cuDNN install on Linux. tried a bunch of things but haven't got it working. I'll ave to stick to standard models for now I guess. Not a DeepFaceLab user because of all the finickiy stuff I'd have to work out to get it to work on linux. You;d think the dev would cater to linux users but eh. whatev. If these new models become standard I'll probably bow out of the community because I can't compete with a p2000 4GB and no working deepfacelab install for linux.

There's a linux fork if you're interested for DFL - https://github.com/lbfs/DeepFaceLab_Linux

andy_ger said:
emsbabs said:
iperov said:
replace DFL folder in _internal

Sorry that I'm bad at understanding. But on your Github I im not sure what to download, so I'll get the updated version.?
Is it the prebuild windows app or something from Github?

By the way, very good job in DFL. I really admire your work and would love to support a little to your futere RTX 2080 Ti as soon as the paycheck ticks in :)

Hi emsbabs,

you have to install the prebuild windows app once.

After that download the github repository



and then extract the content in the zip to the \_internal\bin\DeepFaceLab directory in your prebuild windows app install dir


You can download the latest version here: https://mega.nz/#F!b9MzCK4B!zEAG9txu7uaRUjXz9PtBqg
Just pay attention to the dates.
 
Top