[Guide] Extraction - A Workflow

[manintan]

No, this is really not how the alignments are used. Even when copying from a previous frame, we recommend just using that as the starting point of detecting so that you get valid alignments for the face, otherwise you are just adding bad data that will confuse the model.

There are times when the mask is not on the face because the person is moving and wasnt properly detected. It would be helpful if moving the initial mask points in the manual frame would flow and carry over to the rest of the thumbs in manual, or a way to just have one mask in a frame say be copied to a specific number of other frames. Im talking about those points we can move around.

[Joe100]

Thank you Torzdf.
Yeah I'm not a clever programmer with python skills, sadly!

[manintan]

Deleted my alignments file but still have the extracted faces. How do I make a new alignments file?

[torzdf]

Deleted my alignments file but still have the extracted faces. How do I make a new alignments file?

Not possible, I'm afraid.

Alignments files need to be generated from the source video.

[ipnesterov]

@Vaca: If you see opencl and CPU options in sort you are using a old faceswap version. Update via "tools" -> "check for updates. Should also be a good bit faster after the update,

Нет. Я тоже не вижу этого!! Я обновил и даже переустановил.
Not. I don't see that either !! I updated and even reinstalled.

[torzdf]

That option is no longer in the sort tool, as it is no longer required.

If you need to force sorting on CPU, then check all of the GPUs under the "Exclude GPU" option

[Logry]

Hi, when I try to use the console mode for the extraction I use this command line :

Code: Select all

python faceswap.py extract -i ~/faceswap/src/old.webm -o ~/faceswap/faces/old

And I receive the following error message :

Code: Select all

Traceback (most recent call last):
  File "faceswap.py", line 5, in <module>
    from lib.cli import args
  File "/home/maxime/faceswap-2.0.0/lib/cli/args.py", line 11, in <module>
    from lib.utils import get_backend
  File "/home/maxime/faceswap-2.0.0/lib/utils.py", line 13, in <module>
    from pathlib import Path
ImportError: No module named pathlib

I want to do it console mode beacause the interface presented below doesn't show up when I start faceswap.

[torzdf]

Are you definitely inside the faceswap environment? Pathlib is an explicit pre-requisite

[Logry]

Are you definitely inside the faceswap environment? Pathlib is an explicit pre-requisite

I opened the shell in the /home/ME/faceswap/ indeed.

[torzdf]

Yeah, that's not being in the environment.

conda activate faceswap

[cedenburn]

Hello,

I have question about the sorting process. If there is another face in the frame with the face you are altering do you remove the frame SO NOT to confuse the computer. If so is the frame then not altered by the tool?

Thank you,
Chris

[torzdf]

Not sure I entirely understand the question.

If you have an extracted face (512px extracted image) and your target face is front and center, but there is also another face in the same image, but they aren't front and center then that is fine, keep it. The other face will be masked out anyway.

[robm19]

Hello
What do I do with alignment landmarks when head/facial features are not in camera view?

Example ... head turns right, right eye and eyebrow are no longer visible to the camera, but I still have the right eye and eyebrow landmark points.

Better detail: landmark points 17 through to 21 inclusive and landmark points 36 to 41 inclusive are present in the alignment even though those facial features are not visible to the camera.

Is it safe to delete them? Do I ignore them and let them "do whatever"? Do I stack them to the edge of the face?

Thank you very much in advance
ps I did search through the forum and did not find what I was looking for - if I missed it, please be kind enough to point me to it.

[bryanlyon]

Hello
What do I do with alignment landmarks when head/facial features are not in camera view?

Example ... head turns right, right eye and eyebrow are no longer visible to the camera, but I still have the right eye and eyebrow landmark points.

Specific landmark points really don't matter. We only use the overall alignment. If the face is properly centered and placed in the image (matching other images in the set) then you're fine to ignore it. If it isn't, you can try to put the landmarks approximately where they belong to correct the alignment or if you're just working on a training set, delete the misaaligned faces.

[robm19]

Thank you! Makes sense

[what]

Tutorial claims that the swapping process uses lots of memory but I'm finding odd behavior. It really doesn't use much system memory, pegs the CPU @ 50% and only uses the first GPU is finds. Is this expected normal behavior?

[torzdf]

Inference is a single GPU task, so that is normal. It will use the amount of CPU it requires to process the data. Memory use is exponential. Lots of images = lots of memory use.

[TomLbj]

thanks you are awesome! I think I have a question!

1: After completing the steps of Clean the Alignments File, can I start training?
2: Manually Fix Alignments It looks like I have rewritten the process of extracting alignment files again? So, when I execute Manually Fix Alignments, is the .fsa file I use just extracted using faceswap.py, or is it the .fsa file after executing the Clean the Alignments File process?

[torzdf]

1: After completing the steps of Clean the Alignments File, can I start training?

Yes. that should be fine

2: Manually Fix Alignments It looks like I have rewritten the process of extracting alignment files again? So, when I execute Manually Fix Alignments, is the .fsa file I use just extracted using faceswap.py, or is it the .fsa file after executing the Clean the Alignments File process?

You will always use the alignments file <video name>_alignments.fsa. This is always the latest alignments file. The version which has numbers at the end (eg <video name>_alignments_202310290000.fsa is a back up version that is always taken whenever you make changes to the alignments file. This is so that you can roll back to an earlier version if you make a mistake.

If you are happy that the latest alignments file is correct, then you can safely delete any version of the alignments files that have numbers at the end.

[Kappanator]

hello, is there an extraction/installation I missed if the detected VRAM is as follows?
12/27/2024 22:30:46 INFO Reset batch sizes due to available VRAM: Detect: 2, Align: 2, Mask: 1

I've selected S3FD (default batch size: 4), Fan (default batch size: 12), Bisenet-Fp (default batch size: 8) for the extraction step and during installation picked Setup for NVIDIA GPU option
Also I have an NVIDIA GeForce RTX 4060 Ti for a GPU.