Retouching Algorithm

This project is an unofficial implementation of FabSoften: face beautification via dynamic skin smoothing, guided feathering, and texture restoration. Also this project is a python module written in C++.

Skin

Eye and hair

Another skin samples

1

2

3

Requirements

• Make sure that you have all libs in /external (they are git submodules):

git submodule update --init --recursive

• Also you should install OpenCV (>= 4.8 — older ONNX importers cannot load the parsing model)
• Neural face-parsing weights (required — this is the face definition: the mask boundary follows the real hairline/hat/glasses/neck). Download the MIT-licensed BiSeNet weights into assets/:

curl -L -o assets/face_parsing.onnx https://github.com/yakhyo/face-parsing/releases/download/weights/resnet18.onnx

Building

mkdir build
cd build
cmake ..
cmake --build . -j

Tests

The C++ test suite (GoogleTest) is built by default. Run it from the build directory:

ctest
# or directly:
./tests/retouching_tests

Disable it with cmake .. -DRETOUCHING_BUILD_TESTS=OFF.

CLI

The native retouch binary runs the whole pipeline — no Python, no Docker. It is built together with everything else (target retouch_cli) and lands right in the repo root as ./retouch:

make                                           # -> ./retouch (finds the build dir itself)

./retouch photo.jpg                            # -> photo_retouched.jpg
./retouch photo.jpg -o smooth.png --debug      # custom output + masks in ./debug/
./retouch photo.jpg --strength 0.7 --max-radius 12   # gentler smoothing
./retouch --help                               # all options

Handy make shortcuts for the edit-build-try loop:

make run IMG=photo.jpg                         # rebuild + retouch with debug dumps
make run IMG=photo.jpg ARGS="--strength 0.7"   # same, extra CLI options
make test                                      # rebuild + run the C++ tests

(make auto-detects cmake-build-debug/build; override with BUILD_DIR=<dir>. The explicit command is cmake --build <build-dir> --target retouch_cli.)

Every smoothing knob is exposed:

Option	Default	Effect
--strength <x>	1.0	ADF eps multiplier: <1 delicate, >1 harder
--eps-cap <x>	300	cap for eps = 5·spots+100 (0 = uncapped)
--max-radius <n>	20	largest smoothing window radius
--texture-decay <x>	0.5	higher = more mid-frequency texture retained
--no-spots	off	disable blemish detection/inpainting
--debug [dir]	./debug	dump per-face skin mask, alpha matte, P_skin map, blemish mask, concealed/smoothed/restored frames

The model is auto-searched (--model, $RETOUCH_PARSING_MODEL, ./face_parsing.onnx, ./assets/face_parsing.onnx, next to the binary).

Docker (alternative)

If you prefer a container (on Linux prefix with sudo if needed):

docker build -t retouching-module .
docker run --rm -v "$PWD:/work" retouching-module photo.png
docker run --rm -e RETOUCH_DEBUG_DIR=/work/debug -v "$PWD:/work" retouching-module photo.png

After Compiling

(Only Windows issue for me, on Linux it works perfectly without this step) Put necessary dll's into the Release folder All necessary dll's you can find in assets or you can manually add it via building and compiling OpenCV

Example of work in python

import smoothingmodule as sm

im = sm.RetouchingImg("Orig1.png", "face_parsing.onnx")

Implementations

• Preprocessing
  • Face Detection (dlib HOG)
  • Binary Skin Mask (neural face parsing, CelebAMask-HQ classes; hair/hat/glasses/neck/eyes/brows/lips excluded)
• Skin Mask Generation and Refinement
  • GMM Clustering (Bayes posterior skin probability)
  • Guided Feathering (soft alpha matte)
  • Blemish Detection & Concealment (DoG → Canny → inpaint)
• Skin Imperfection Smoothing
  • Dynamic Smoothing Hyperparameter (ADF: r = 10·P_skin + 10, eps = 5·num_spots + 100)
• Skin Texture Restoration (DWT)

Functions

Name of the function	Discription	Input
Retouching	Retouching array-like data	Array-like data and Model Dir
RetouchingImg	Retouching photo (jpg, png, e.t.c.)	Photo Dir and Model Dir