Mr DeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
digdeepWhat is the best deepfake app?
#1
When searching online there are so many different apps. Which one is best?
#2
The Python scripts give the best results right now if you can get them to install properly. FakeApp is not as efficient, but it is also the easiest to use at the moment. I plan on writing a tutorial of each tool soon.
#3
If you are good with python scripting then it would be very good if you could edit the fakeapp 1.1.0 version Tongue i use that program and its for me easy to use it. Most of the people dont have idea how to use the python version. But i think its not so easy to edit the program maybe.
#4
(07-22-2018, 01:24 AM)Nihlus Wrote: You are not allowed to view links. Register or Login to view.If you are good with python scripting then it would be very good if you could edit the fakeapp 1.1.0 version Tongue i use that program and its for me easy to use it. Most of the people dont have idea how to use the python version. But i think its not so easy to edit the program maybe.

A lot of people still use the original fakeapp 1.1, it still gives pretty good results, and is friendlier with less powerful GPUs I believe.
#5
so i'm running myfakeapp with an AMD setup, but my GPU doesn't want to participate (not going to do the Linux setup that one tutorial had..too much effort).  anyone have any tips to improve training efficiency?  running a batch of 64 on my cpu using the original trainer.  so it's just taking forever.  anyone have a recommended batch size?  what about training the B face against itself for a little bit?  does that help with loss values?  lowest I've gotten so far is a loss of about .039.  that was after about 24 hours of training.  if the only answer is "you just need to give it more time" then fine.  I was just hoping there was a more efficient way than just brute force.  I'm ok with just using brute force if that's the only way.
#6
(07-25-2018, 11:52 PM)wifepics Wrote: You are not allowed to view links. Register or Login to view.so i'm running myfakeapp with an AMD setup, but my GPU doesn't want to participate (not going to do the Linux setup that one tutorial had..too much effort).  anyone have any tips to improve training efficiency?  running a batch of 64 on my cpu using the original trainer.  so it's just taking forever.  anyone have a recommended batch size?  what about training the B face against itself for a little bit?  does that help with loss values?  lowest I've gotten so far is a loss of about .039.  that was after about 24 hours of training.  if the only answer is "you just need to give it more time" then fine.  I was just hoping there was a more efficient way than just brute force.  I'm ok with just using brute force if that's the only way.

What kind of GPU do you have? If you have a less powerful GPU, it might be better to train on the lowmem model? But then you'd have to start over. I am training on HighRes model now, and it's 30 hours and my loss is still 0.025. It might take me 2 more days for this training to get done.

Once you have a model pre-trained, it'll be faster to train different ones. But I don't think training the same faces against each other is better. How many images are in your face sets? The less the numbers, the faster to train.
#7
ok. that gives me some insight. I got some lower end stuff. AMD Radeon R5 Graphics with an AMD A8-7410APU. so I guess I just need to set it to train at the start of the weekend. I may try the lowmem setting. It's training. I've done some test vids (really crappy) just to make sure the process works. so I guess it's just giving a lot of time to train a model. thanks!

oh, and I have about 200 pics for face B and running batch size of 64. just switched to 32 to maybe give more change for it to learn
#8
(07-26-2018, 12:59 AM)wifepics Wrote: You are not allowed to view links. Register or Login to view.ok.  that gives me some insight. I got some lower end stuff.  AMD Radeon R5 Graphics with an AMD A8-7410APU. so I guess I just need to set it to train at the start of the weekend.  I may try the lowmem setting.  It's training.  I've done some test vids (really crappy) just to make sure the process works.  so I guess it's just giving a lot of time to train a model.  thanks!

oh, and I have about 200 pics for face B and running batch size of 64.  just switched to 32 to maybe give more change for it to learn

Generally the higher that batch you can get away with, the better. I am surprised you can run fakeapp at all with that =P.

Yes testing is key, I have wasted many hours just training a failed model (I still do this actually lol)
#9
i'm not running fakeapp….i'm running an alternative "myfakeapp". lol. it's made for peons like me.
#10
(07-25-2018, 04:56 AM)dpfks Wrote: You are not allowed to view links. Register or Login to view.
(07-22-2018, 01:24 AM)Nihlus Wrote: You are not allowed to view links. Register or Login to view.If you are good with python scripting then it would be very good if you could edit the fakeapp 1.1.0 version Tongue i use that program and its for me easy to use it. Most of the people dont have idea how to use the python version. But i think its not so easy to edit the program maybe.

A lot of people still use the original fakeapp 1.1, it still gives pretty good results, and is friendlier with less powerful GPUs I believe.

I use a i7-5960X and 2xGTX980ti Tongue
Most i can see that the results are better with other app. But i like the fakeapp1.1 because its really easy to use.

Forum Jump:

Users browsing this thread: 2 Guest(s)