search

pozitronik / sinner

Sinner: Sinner Is NoN Exactly Roop
GNU Affero General Public License v3.0
77 stars 9 forks source link
deepfake deepfakes face-enhancement image-processing video-processing

readme

Build Status

sinner: sinner is non exactly roop

Deepfakes and more.

What is it?

This is the rework of the s0md3v/roop that I'm working on for entertainment and educational purposes. It doesn't aim to be popular; it's just a fork made the way I want it to be. The tasks that I aim to accomplish here are:

How do I install it?

The basic installation instructions for now are the same as those in the s0md3v/roop, check them out. In short, you need to install python 3.10 or a later version, VC runtimes (on windows), and desired Execution Provider kit (depending on your hardware and OS). Then you need to install required python packages, and there are some differences in the process:

I have PC with Windows/Linux and no CUDA GPU

Run pip install -r requirements.txt. It will install packages, just enough to run magic on CPU only.

I have PC with Windows/Linux and GPU with CUDA support

Run pip install -r requirements-pc-cuda.txt. It will install packages with CUDA support. Do not forget: you also have to install CUDA drivers as well.

I have x86 Mac

Run pip install -r requirements-mac-x86.txt. It will use only CPU powers, but it should work.

I have Apple Silicon Mac

Run pip install -r requirements-mac-arm64.txt. There is no CUDA, obviously, but there's some hardware acceleration too.

Anyway, packages should be installed successfully. Otherwise, get a look to the command output, usually you may fix minor issues (like version requirements change) by yourself.

If nothing helps, feel free to create an issue with your problem, we will try to figure it out together.

How do I use it?

Go to application folder and run python sin.py with desired set of command-line parameters (or just pick one of examples and make changes to suit your need).

You can get the list of all available command-line parameters by running the program with --h or --help keys. Those commands will list all configurable modules and their supported parameters.

Some modules may have the same parameters. It is okay: those parameters (and its values) are shared. It is also okay, if parameters expected values are different between modules: usually, they will be harmonized in runtime. But if something goes wrong, you will get an explicit error message.

Also, you can read about modules parameters here

Built-in frame processors

There are modules named frame processors, and each processor can perform its own type of magic. You need to choose which frame processor (or processors) you want to use, and provide them with some sources to work on. Here is the list of built-in frame processors:

Take video stream from a video file, swap and enhance all faces to provided source face using CUDA device (e.g. nvidia GPU) and create the new video stream and show it in the preview window.

Command line usage examples

python sin.py --source="d:\pictures\cool_photo.jpg" --target="d:\pictures\other_cool_photo.jpg" --frame-processor=FaceSwapper

Swap one face on the d:\pictures\other_cool_photo.jpg picture to face from the d:\pictures\cool_photo.jpg picture and save resulting image to d:\pictures\cool_photo_other_cool_photo.png (autogenerated name).

python sin.py --source="d:\pictures\cool_photo.jpg" --target="d:\videos\not_a_porn.mp4" --frame-processor FaceSwapper FaceEnhancer --output="d:\results\result.mp4" --many-faces --execution-provider=cuda

Swap all faces on the d:\videos\not_a_porn.mp4 video file to the face from d:\pictures\cool_photo.jpg and enhance all faces quality, both processing will be made using the cuda provider, and result will be saved to the d:\results\result.mp4.

python sin.py --source="d:\pictures\any_picture.jpg" --target="d:\pictures\pngs_dir" --output="d:\pictures\pngs_dir\enhanced" --frame-processor=FaceEnhancer --many-faces --max-memory=24 --execution-provider=cuda --execution-threads=8

Enhance all faces in every PNG file in the d:\pictures\pngs_dir directory using the cuda provider and 8 simultaneous execution threads, with limit of 24 Gb RAM, and save every enhanced image to the d:\pictures\pngs_dir\enhanced directory.

Real-time player

This feature is still in the alpha stage, so things can be changed. There's not much to document yet; it's better to try it yourself, by running it:

python sin.py --gui

FaceSwapper demo

Virtual camera feature

You can use sinner to create a virtual real-time face-swapped camera, and use it with your software. This feature is managed by the WebCam module. You can refer to the module's live help (python sin.py -h) to find information about currently supported features. Here is a description of the common setup process:

First, you will need a real camera for the actual input. If you have only one camera, it will be used by default. Otherwise, you should pass the camera index as the --input parameter. You can also pass a path to an image or a video file to use them as input sources, although it may not be as enjoyable.

Second, you will need to create a virtual camera. The currently supported devices are suggested in the help output for --device key. If there are no supported options, install OBS Studio to obtain the obs virtual camera device. Other virtual camera software may be used, but OBS is recommended.

Third, you will need to set up the processing chain as usual.This involves selecting an image with the source face, configuring processing modules, and so on. Here are some examples:

python sin.py --camera --device=obs --source="path\to\source\face.jpg" --frame-processor=FaceSwapper --many-faces --execution-provider=cuda --preview

Capture a video stream from a webcam, apply face swapping to replace all faces with a provided source face using a CUDA device (e.g., an NVIDIA GPU), and generate a new video stream using the OBS virtual camera. Additionally, display the stream in the preview window.

Since real-time processing can be slow, depending on hardware capabilities, it might be a good idea to reduce the resolution of processed frames to achieve a higher frame rate:

python sin.py --camera --device=obs --source="path\to\source\face.jpg" --frame-processor=FaceSwapper --many-faces --execution-provider=cuda --width=480 --height=320

There is no need to use FrameResizer module in this case, --width and --height keys will be handled by the WebCam module itself.

You also can use the camera feature to obtain a real-time preview of a video file:

python sin.py --camera --device=no --input="path\to\video.mp4" --source="path\to\source\face.jpg" --frame-processor FaceSwapper FaceEnhancer --many-faces --execution-provider=cuda --preview

Configuration file

You can store commonly used options in the configuration file, to make them apply on every run by default. Just edit sinner.ini file in the application directory and add desired parameters inside the [sinner] section as key-value pairs.

Example:

[sinner]
keep-frames = 1
many-faces = 1
execution-provider = cuda
execution-threads = 2

It is also possible to configure modules separately this way. Just create/modify a config section with the module name, and all key-value pairs from this section will be applied only to that module.

Example:

[sinner]
execution-threads = 2

[FaceSwapper]
execution-threads = 4

In the example above FaceSwapper will run in four execution threads, when other modules will run in two threads (if they support this parameter). Module configurations have the priority over global parameters (even if they passed directly from the command line).

Any parameter set from command line will override corresponding global (not module) parameter from the ini file.

You also can pass path to the custom configuration file as a command line parameter:

python sin.py --ini="d:\path\custom.ini"

How to handle output videos quality/encoding speed/etc?

In brief, sinner relies on the ffmpeg software almost every time video processing is required, and it's possible to utilize all the incredible powers of ffmpeg. Use the --ffmpeg_resulting_parameters key to control how ffmpeg will encode the output video: simply pass the usual ffmpeg parameters as the value for this key (remember not to forget enclosing the value string in commas). There are some examples:

And so on. As you can find, there are a lot of different presets and options for the every ffmpeg encoder, and you can rely on the documentation to achieve desired results.

In case, when ffmpeg is not available in your system, sinner will gracefully degrade to CV2 library possibilities. In that case all video processing features should work, but in a very basic way: only with the software x264 encoder, which is slow and thriftless.

FAQ

:question: What are the differences between sinner and roop?
:exclamation: As said before, sinner has started as a fork of roop. They share similar ideas, but they differ in the ways how those ideas should be implemented. sinner uses the same ML libraries to perform its magic, but handles them in its own way. From a developer's perspective, it has a better architecture (OOP instead of functional approach), stricter types handling and more comprehensive tests. From the point of view of a user, sinner offers additional features that Roop currently lacks.

Also, roop is dead, baby, roop is dead.

:question: Is there a NSWF filter?
:exclamation: Nope. I don't care if you do nasty things with sinner, it's your responsibility. And sinner is just a neutral tool, like a hammer or a knife.

:question: Can I use several execution providers simultaneously?
:exclamation: You can try. Seriously, you can set --execution-provider cuda cpu, and look, what will happen. May be it will work faster, may be it won't work at all. It is a large space for experiments.

Credits

License

GNU GPL 3.0