The shaders implement chroma upscaling based on the closed least squares solution for linear regression, inspired by the adoption of the same technique in modern video codecs.
Since a simple linear regression obviously doesn't take into account pixel distance, the prediction is mixed with the output of a normal resampling filter based on how high the correlation between luma and chroma is. Additionally, it's also smoothed with a bilateral filter.
You can control which local linear regressions you want to use with USE_12_TAP_REGRESSION and USE_8_TAP_REGRESSIONS. You can also configure the smoothing step with the distance_coeff and intensity_coeff inputs.
The lite variant offers slightly better performance at the expense of some quality.
Add something like this to your mpv config:
vo=gpu-next
glsl-shader="path/to/shader/CfL_Prediction.glsl"