Corridor Digital's Faceswap videos.

Display post links to interesting swaps
Post Reply
User avatar
bryanlyon
Site Admin
Posts: 464
Joined: Fri Jul 12, 2019 12:49 am
Answers: 39
Location: San Francisco
Has thanked: 3 times
Been thanked: 112 times
Contact:

Corridor Digital's Faceswap videos.

Post by bryanlyon »

Good examples of what can be done with FS + some post processing. Also a cool Behind The Scenes of the action. Check it out.


User avatar
Sanscouer
Posts: 2
Joined: Thu Jul 25, 2019 4:13 pm

Re: Corridor Digital's Faceswap videos.

Post by Sanscouer »

They do a nice work on that using the DeepFaceLab, they have other with making a Tom Cruise


User avatar
bryanlyon
Site Admin
Posts: 464
Joined: Fri Jul 12, 2019 12:49 am
Answers: 39
Location: San Francisco
Has thanked: 3 times
Been thanked: 112 times
Contact:

Re: Corridor Digital's Faceswap videos.

Post by bryanlyon »

Sanscouer wrote: Thu Jul 25, 2019 4:15 pm

They do a nice work on that using the DeepFaceLab,

This was actually almost entirely made with FaceSwap, not DFL. Some shots were DFL, but the vast majority were FaceSwap.

In fact, check out the second video where they explain exactly that.


User avatar
AndrewB
Posts: 8
Joined: Tue Nov 12, 2019 10:16 am
Has thanked: 1 time

Re: Corridor Digital's Faceswap videos.

Post by AndrewB »

Did they used the Villain model for most of the shots? The final result is great!


User avatar
bryanlyon
Site Admin
Posts: 464
Joined: Fri Jul 12, 2019 12:49 am
Answers: 39
Location: San Francisco
Has thanked: 3 times
Been thanked: 112 times
Contact:

Re: Corridor Digital's Faceswap videos.

Post by bryanlyon »

A quote from Niko about the stuff he was using.

nikopueringer 07/14/2019
Answering a couple of the questions from further up;
We used both FS and DFL. In FS we used the villain model, with 100% coverage and extended mask. I first pretrained the villain model for 50k iterations on random faces, then began training it on Keanu from there. The face tone and shading was more natural and “stable” with my pretraining vs no pretraining, but my method then was still rough, and there are things I would try differently if I were to go back.
What you see in the video is almost always the villain model, with the DFL mode being used for the closeup at the end and two other collide-ups in the video
The villain model was trained to 200k iterations on top of the 50k pretrain
It looked more faithful to Keanu’s likeness, but occasionally struggled to match the lighting, and was limited to 128x128 resolution.
The DFL model was SAE, but using the LIAE option, so similar to using IAE in fsceswap.
That was pretrained as well, and then ran on Keanu to 250k iterations. The resolution was 256x256, so noticeably clearer.
The DFL model incorporated more of the base face when generating the new face, so the lighting matched better and some more complicated expressions came through clearer. The downside is that it looked close to Keanu, but sometimes wasn’t completely faithful to his identity.
The ending “selfie” moment is DFL.
So both models had their perks. We focused on using Villain in the beginning to help break the barrier of disbelief, and once that was achieved we had more flexibility in using the DFL model when needed.
Typed this on my phone so there are a few auto-correct errors...


User avatar
Now
Posts: 10
Joined: Sat Jan 04, 2020 6:30 pm
Has thanked: 6 times
Been thanked: 1 time

Re: Corridor Digital's Faceswap videos.

Post by Now »

these are bloody brilliant..... thanks for posting..... would love some tips to get it to this level of brilliance and realism.


User avatar
Now
Posts: 10
Joined: Sat Jan 04, 2020 6:30 pm
Has thanked: 6 times
Been thanked: 1 time

Re: Corridor Digital's Faceswap videos.

Post by Now »

got a little excited.. the tips are provided by Mr Lyon, thanks ! Would love to know what the following means :

  1. "with 100% coverage and extended mask"

  2. "pretrained the villain model for 50k iterations on random faces" why would he do this...or did he mean random faces of Keaunu ?

  3. I found the villian option but not sure where the DFL mode is (im guessing that may be a mode in another software , not faceswap)
    brilliant, thanks for forwarding his summary,

  4. Also , would i be right in saying the 256x256 size is better for close up shots... and the smaller size is better for the further away shots?


User avatar
bryanlyon
Site Admin
Posts: 464
Joined: Fri Jul 12, 2019 12:49 am
Answers: 39
Location: San Francisco
Has thanked: 3 times
Been thanked: 112 times
Contact:

Re: Corridor Digital's Faceswap videos.

Post by bryanlyon »

Now wrote: Wed Jan 08, 2020 4:57 pm
  1. "with 100% coverage and extended mask"

See viewtopic.php?f=6&t=146#config

Now wrote: Wed Jan 08, 2020 4:57 pm
  1. "pretrained the villain model for 50k iterations on random faces" why would he do this...or did he mean random faces of Keaunu ?

There is some debate as to whether pretrain is useful or not. Pretraining may be able to help when there is a limited amount of data.

Now wrote: Wed Jan 08, 2020 4:57 pm
  1. I found the villian option but not sure where the DFL mode is (im guessing that may be a mode in another software , not faceswap)
    brilliant, thanks for forwarding his summary,

The DFL here is referring to another software, yes. But the model they used is now in Faceswap as "DFL-SAE".

Now wrote: Wed Jan 08, 2020 4:57 pm
  1. Also , would i be right in saying the 256x256 size is better for close up shots... and the smaller size is better for the further away shots?

Best to train at the highest quality you think you might need and your GPU can handle. Usually this will be 128 for most models. It's important to note that this is entirely separate from the extract resolution which defaults to 256 and probably shouldn't be changed without a good reason.


User avatar
Now
Posts: 10
Joined: Sat Jan 04, 2020 6:30 pm
Has thanked: 6 times
Been thanked: 1 time

Re: Corridor Digital's Faceswap videos.

Post by Now »

Thanks Bryan..... gonna give it another go.... appreciate the advice ..


Post Reply