0
13

[–] deepfaps 0 points 13 points (+13|-0) ago 

Please upvote this reply so I can get enough CCP to post. I have a good one ready to upload. Thanks.

0
6

[–] FeepDakeNoob 0 points 6 points (+6|-0) ago 

So what is the gold standard 'best set-up' with this software right now?

I started messing around with fakeapp 2.1 then switched to 1.1 which seemed more flexible/powerful and got some more decent results.

Is it worth looking at 2.2 or moving to the python scripts?

Also, it's it explained anywhere what changing the values for nodes, layers and batch size will have on performance vs quality?

0
0

[–] frosty3907 0 points 0 points (+0|-0) ago  (edited ago)

Would love to know this also, haven't seen any results form the alt scripts (GAN + faceswap) for comparison to know if they're worth the effort aside from the examples on github.

e: Also still using 1.1 here, haven't tried 2.2 due to all the crappery since 1.1

0
0

[–] FeepDakeNoob 0 points 0 points (+0|-0) ago 

I started with 2.1, then went to 1.1 once I understood things a bit more.

I have not been able to get 2.2 working at all, and I'm not trying all that hard because it seems problematic.

0
4

[–] supervisedlearner 0 points 4 points (+4|-0) ago 

How interchangeable are the generated models? If I started training with fakeapp, can I continue with the same model with myfakeapp or deepfakes/faceswap?

0
0

[–] recziat [S] 0 points 0 points (+0|-0) ago 

It depends on what FakeApp uses internally but it's extremely likely that the models will be incompatible. With the python scripts, you have different types of models to choose from, including the default, classic model, which might be compatible. Apart from that, they shouldn't be interchangeable at all.

0
0

[–] supervisedlearner 0 points 0 points (+0|-0) ago 

Is there any documentation on these different models? Like, what is lowmem, GAN, what does it mean to allow growth, what is an epoch

0
2

[–] justSomeAsshole 0 points 2 points (+2|-0) ago 

deepfakes/faceswap has a GUI, you just run it with the argument gui

0
0

[–] recziat [S] 0 points 0 points (+0|-0) ago 

Ah, they finally hacked one together, great! And it doesn't even look that bad.

0
2

[–] derram 0 points 2 points (+2|-0) ago 

0
1

[–] deepshakes 0 points 1 points (+1|-0) ago 

I posted this on /v/facesets but I figured I'll add it here as well:

I don't think the creator of the script (MotorCityCobra) has a voat account, but his script face-ripper-9000, based on the GetFaces script, will greatly simplify your faceset making process compared to using the FakeApp extract feature manually. The reason I'd reccommend this script over the original script is that GetFaces can only handle one video at a time, will output your facesets with incorrect dimensions, and will name output filenames randomly making sorting frustrating.

Currently the faceset making process consists of downloading a set of videos, converting them each to individual frames using FFMPEG, searching through each data set to delete any frames with multiple faces, extracting your results and then going back through your data to delete any errors. With face-ripper-9000, all you need to do is download a set of videos as well as a picture of the person who's face you want to extract, and the script will run through each clip and extract only frames containing faces that match the input photo, outputting them in 256x256, ready to be used for training. It will also automatically put any particularly bad frames into a seperate folder.

You'll still have to go through your final set and check for any errors, but it completely removes the hassle of having to first convert your videos to individual frames, and having to delete multiple faces in your data set before extracting them.

0
0

[–] FeepDakeNoob 0 points 0 points (+0|-0) ago 

Interesting. I'm going to give this a shot.

0
0

[–] m0rm0 0 points 0 points (+0|-0) ago  (edited ago)

I'm stuck on a step while following the tutorial linked on this subverse (https://voat.co/v/DeepFake/2388368), the DeepFakes FakeApp Tutorial 1.1. I would have made this as my own thread and would have submitted the two aligned image sets I created below on the FaceSets forum, but I believe my account is too new. Regardless I still have valuable sets below if anyone else would like to repost them because I can't... anyway:

http://archive.is/fiO8v#selection-143.0-146.0

Here is a screenshot of what happens when I try to do the "Begin Training" step outline above: http://imgbox.com/ZehyiX1S

So I'm unable proceed and not really sure how to diagnose why.

This is my first attempt and I'm trying to combine the ASMRist "Springbok" with the pornstar Nancey aka Miriama Kunkelova (it may come out poorly, but at least I will have tried and learned. And I want nothing more than to see a good video of Springbok, lol). Furthermore, I will upload both sets to help the community as well, maybe someone else will be able to put it together too and maybe someone can think of a better match for their faces. Here are the aligned images for each woman, maybe they need more images for it to be perfect but there are many:

Springbok ASMR (7116 aligned images): https://www44.zippyshare.com/v/VLKLHrTq/file.html

Miriama Kunkelova (11737 aligned images): https://www11.zippyshare.com/v/30YGd5nz/file.html

Well, thanks for any and all help! I wanted to post this on the FaceSets subverse too but it won't let me post in the Facesets tutorial area whatsoever. So please let me know if anyone can assist or wants to help me use these facesets I've made to make it work.

0
0

[–] turbowatt 0 points 0 points (+0|-0) ago  (edited ago)

I made a dedicated post about this, but it belongs in this topic too:

Many people here may not be familiar with how FakeApp works. It is basically a GUI wrapper around some Python scripts. If you're a developer interested in getting into the project, or you don't want to make videos with watermarks, you might want to try using faceswap, the modern version of the Python scripts.

A lot of people don't use the Python scripts directly since some of the dependencies (mainly dlib) are a pain to install. Well, you don't have to. A couple months ago, DeepFakesClub made portable versions of faceswap, which is the main deepfake repo, and dfaker. It doesn't include the most recent version of faceswap, which has GUI support. However, the way it works is that there's a portable version of WinPython with the dependencies installed. The faceswap .py files are sitting in a specific directory, python-3.6.3.amd64/faceswap. You can drag and drop in the latest version of faceswap, or even mess around with the Python files if you like.

You may want to mention this tool in the main post! Along with the fact that you can easily update the scripts by downloading the new .py files and dropping them into the right directory.

Here are the original posts. They all link to the same mega.nz link with the WinPython download:

http://archive.is/O64Mb

https://github.com/deepfakes/faceswap-playground/issues/86

Some usage info is also mentioned on the DeepFakeClub forums, but apparently linking to them is banned for some reason. DeepFakeClub doesn't appear in the banned sites list, though.

load more comments ▼ (8 remaining)