I haven't been able to test this much, but outside of spatial constrains, this seems to work decently well, so it should be useful for LUT generation already. Outputting images is too slow for it to be of much interest, anyway.
Use moving least squares to generate color transferring 3D LUTs or transferred images.
Based on: Hwang, Y., Lee, J. Y., Kweon, I. S., & Kim, S. J. (2019). Probabilistic moving least squares with spatial constraints for nonlinear color transfer between images. Computer Vision and Image Understanding, 180, 1-12. https://doi.org/10.1016/j.cviu.2018.11.001
- No probabilistic modeling is performed, as no image registration is necessary for our use case.
- Extrapolation just checks if nearest neighbor is within a sphere around the bin center. This is faster and should be good enough.
- No GPU processing (yet?). (Mine's too old.) I hope this is the reason processing is way slower than the paper states.
usage: MLSColorTransfer [-S] [-P] [-s SPATIAL-WEIGHT]
[-r COLOR-WEIGHT] [-p PREFILTER-STRENGTH]
[-c CONTROL-POINTS] [-l LUT-SIZE] [-i IMAGE]
[-h] source target filename
Moving least squares for color transfer.
positional arguments:
source Source image(s). For multiple images, this
should be a folder with shared filenaming
between source and target images.
target Target image(s). For multiple images, this
should be a folder.with shared filenaming
between source and target images.
filename Output image (png) or 3D LUT (cube) file name.
optional arguments:
-S, --spatial Enable using spatial info. Can only be used if
outputting an image. Experimental.
-P, --no-prefilter Disable prefilter.
-s, --spatial-weight SPATIAL-WEIGHT
Spatial domain weighting strength. (type:
Float64, default: 20.0)
-r, --color-weight COLOR-WEIGHT
Color space domain weighting strength. (type:
Float64, default: 20.0)
-p, --prefilter-strength PREFILTER-STRENGTH
Prefilter strength. (type: Float64, default:
2.0)
-c, --control-points CONTROL-POINTS
Number of control points to use. Lowering this
will lead to worse results and potential
crashes, but will decrease computation time.
It's recommended to set to 1% of total pixel
count for best results. (type: Int64, default:
1000)
-l, --lut-size LUT-SIZE
3D LUT size. (type: Int64, default: 33)
-i, --image IMAGE Image to apply transfer to. Must be set if
outputting an image.
-h, --help show this help message and exit
Some things can use multi threading. You can pass --julia-args -t auto to take full advantage of this. However, you might run out of memory.
Images need to be cropped and resized to match more or less exactly. Misaligned images will lead to terrible results.
Spatial information doesn't really work yet.