hzrJBGPpgN

Take a video and replace the face in it with a face of your choice. You only need one image of the desired face. No dataset, no training.

Also allow row render (original, faceswap, enchance face via codeformer or gfpgan):

colab link

easy installation guide for windows users

1) install visual studio 2022 with desktop development C++ and python development (not sure about python development)
2) install python 3.10.x (any 3.10)
3) download the last version of roop
4) pip install virtualenv
5) virtualenv venv
6) start venv\scripts\activate.bat
7) pip install -r requirements.txt
that's for cpu (sometimes works for gpu for some reason)
if you want nvidia gpu to work:
don't go for cpu
1) install visual studio 2022 with desktop development C++ and python development (not sure about python development)
2) install cuda 11.7 (https://developer.nvidia.com/cuda-11-7-0-download-archive)
3) download cudnn 8.9.1 for cuda 11.x https://developer.nvidia.com/rdp/cudnn-archive
4) unpack cudnn over C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7 with replacement
5) install python 3.10.x (any 3.10)
6) download the last version of roop
7) pip install virtualenv
8) virtualenv venv
9) start venv\scripts\activate.bat
10) pip install torch torchvision torchaudio --force-reinstall --index-url https://download.pytorch.org/whl/cu117
11) pip install -r requirements.txt
and yes, don't forget to download ffmpeg https://ffmpeg.org/download.html
and inswrapper_128.onnx https://drive.google.com/file/d/1eu60OrRtn4WhKrzM4mQv4F3rIuyUXqfl/view?usp=drive_link

usage example

#usage is simple:
python run.py --gpu --gpu-threads %number_of_threads%

for number of threads I recommend to play, for nice approximation of first step is: amount of threads = (GPU VRAM - 1)/800

Face enchancement

You can also apply face enchancement by codeformer or gfpgan (slow!!)

For gfpgan (slightly faster then codeformer)

To do:

In options/core.json change "default_chain": "facedetect,faceswap,gfpgan"

Or if you want just enhance your video:

In options/core.json change "default_chain": "facedetect,gfpgan"

For codeformer

To do:

Download plugin here: https://github.com/janvarev/chain-img-plugin-codeformer
Copy it to plugins folder
Install requirements.txt
In options/core.json change "default_chain": "facedetect,faceswap,codeformer"

Or if you want just enhance your video:

In options/core.json change "default_chain": "codeformer"

It's recommended to set settings in options/codeformer.json like:

{
    "background_enhance": false,
    "codeformer_fidelity": 0.5,
    "face_upsample": true,
    "skip_if_no_face": true,
    "upscale": 1
}

For gfpganonnx

Effective implementation of GFPGAN on ONNX (I gain 2.5x speedup)

To use:

convert GFPGAN model to ONNX format https://github.com/xuanandsix/GFPGAN-onnxruntime-demo/tree/main
your resulted model must be placed as ./models/GFPGANv1.3.onnx
In options/core.json change "default_chain": "facedetect,faceswap,gfpganonnx"
Note: If you wanna just ehchance face on current video, use "default_chain": "facedetect,gfpganonnx"

Faceswap selective

Allow you to selective swap faces on img/video, the same as in "refacer" project

To use:

run program at least once
In options/core.json change "default_chain": "facedetect,faceswap_selective" (you can use enchancer if you want)
In options/plugin_faceswap_selective.json change "selective" to "char1.jpg->new1.jpg||char2.jpg->new2.jpg", where
- char1.jpg is path to file with someone face from original video
- new1.jpg is path to file with face that replace char1
- -> is separator
- || is separator for groups (you can set char2,new2 etc.)
- NOTE: face in roop interface will not being used in this case
Each face on scene will be compared to reference in "selective" option. If distance is lower than "max_distance" - face will be replaced (try to adjust "max_distance" if you have problems).

Special thanks

https://github.com/janvarev/chain-img-processor licensed under MIT

paolo1234 / Richard-roop