hi, I have tried your code, but seems a bit wrong here =>
color_transfer / color_transfer / init.py

in the 42~45 lines,

scale by the standard deviations

l = (lStdTar / lStdSrc) * l
a = (aStdTar / aStdSrc) * a
b = (bStdTar / bStdSrc) * b

the StdSrc and StdTar should be replaced with each other.

I will try more test images and feedback to you~

Line 26
target = cv2.cvtColor(target, cv2.COLOR_BGR2LAB).astype("float32")

while trying to use differents jpg, raise error during the conversion to float32:

target = cv2.cvtColor(target, cv2.COLOR_BGR2LAB).astype("float32")
cv2.error: ........\opencv\modules\imgproc\src\color.cpp:4118: error: (-215) (scn == 3 || scn == 4) && (depth == CV_8U || depth == CV_32F) in function cv::cvtColor

When I use following line to install the lib,
pip install color_transfer
and use codes in the example.py, there will be errors:
TypeError: color_transfer() got an unexpected keyword argument 'preserve_paper'
TypeError: color_transfer() got an unexpected keyword argument 'clip'

so does the files in the pip lib older than the current repo?

I believe the original paper means that we want the source image to mimic the color performance of the target image, and output a transferred source image. However, in your implementation, the target image is instead transferred. This is what makes the following formulas inconsistent with the paper.

I have identified a simple enhancement of the basic Reinhard method which for some input image pairs can produce images of far better quality.

https://github.com/TJCoding/Enhanced-Image-Colour-Transfer

The implementation is in C++ only for the moment, but an executable is provided for those who wish to try out the new method. Also a web app is now available which permits a comparison of various processing methods.
https://www.dustfreesolutions.com/CT/CT.html

I have an adaption of your code which implements the enhanced method in Python. How would you like me to deal with this? I could take a branch and issue a pull request or I could simply post it in my repository in a directory 'Python implementation of Enhanced-Image-Colour-Transfer' and fully credit you for your contribution. Although I have used your code as a template, there is quite a wide divergence from the original so a merge might not be so simple. The choice is yours.

The images below show target image, source image, output image from current processing, output image from enhanced processing.

Any interest in including a utility function for choosing the 'truest to source' color transfer?

I've implemented an (unoptimized) auto_color_transfer function that attempts to choose the best combination of clip & preserve_paper by comparing each potential result to the source image in the HSV color space using chi2 distance. The mean absolute error is calculated between the channel means of the source & candidate images. Since the 'truest' color transfer isn't always the most aesthetically pleasing the function also returns a comparison matrix image (see below).

In this case, the bottom left image is chosen as the truest transfer by the current method.

EDIT: updated distance used to determine 'truest' transfer

Hi. Thanks for the code! Impressive work!
I find that for some images, the code seems to produce the wrong output. Here is the example:
source:

target:

output:

I will appreciate it if you can offer me some guidance. Thanks ahead of time!