Maybe an applied example of what CFA drizzle can be good for, and how I use it, might be of some use.
Here is a side-by-side example using data which ultimately ended up producing
this photograph of Andromeda, showing, at this zoomed out scale, some of the reasons why CFA drizzle can be useful (e.g. for color accuracy), and
zoomed in 1:1 showing some of the smaller-scale structural changes it affects.
Sorry about the 100% scaling, folks. Not sure how to limit it here except with a small image.
Click here for 1:1 scale (carefully compressed to avoid much loss of color and detail)
Of note, for a CFA drizzle, I would just use a drop shrink of 1. You will still get some enhancement of detail regardless, but a drop shrink of 1 is sufficient to bypass debayer interpolation. And I would rather deal with any enhancement of detail beyond this point in post-processing (especially simple as a mild application of BlurX structure detail enhancement). Those and other linear post-processing can be applied ahead of Integer Resample as one option if a larger resolution isn’t desired.
CFA drizzle integration is, in my opinion, a valuable tool when working with CFA data. And even if I don’t use it to upscale, I tend to appreciate it solely for the ability to bypass debayer interpolation. It can help a lot with certain types of aberrations (especially in stars), helps with some forms of color noise, balances the noise profile (a non-CFA drizzle of color data tends to create noise which varies in peaks and valleys that I find harder to work with and process around), and improves accuracy of color calibration.
The applied difference can vary quite a bit depending on the image. With a larger amount of well-dithered exposures differences can become less pronounced, and some of the color calibration and star form/aberration benefits are most pronounced when working with undersampled data. But I’ve certainly tested it a mess o’ times over trying to decide what workflow is best for my own images, at different scales, trying to sort out which is best for me.
Usually the S/N reading for the CFA-drizzled image is lower, and you can usually see that the non-CFA drizzled image will auto-stretch a bit brighter in mid-tones. But I’m not sure I take that strictly at face value as when you stretch the two samples to match and use noise reduction, as results tend to be a bit more comparable than the readout number implies. The noise profile of the CFA-drizzled image tends to be easier to work with in many ways. You tend to get some sort of vestige of the CFA showing up around stars in a CFA drizzle (you can see a much less dramatic example of it here), but if you look at the non-CFA drizzled image you get somewhat block-like artifacts in their place. Noise reduction can mitigate pixel-level artifacts as in this example reasonably well.
When CFA data is drizzled in WBPP it applies a CFA drizzle. Need to run Drizzle Integration independently to turn off the CFA drizzle option. And personally, I’m not sure how useful that option has been. I have found some uses for it with color data, but mostly I prefer the CFA drizzle result (for the reasons why it exists).
In the times I have been able to lend a hand to someone, complaints like “SPCC does not give me a good color calibration,” when working with color sensor data, tend to be helped out a lot by using a CFA drizzle (if it’s not a different issue like pedestals not being added when needed).
I don’t think I’ve ever seen a pattern quite as pronounced as OP’s, though. This dataset could have been better suited for a 2x drizzle with more subs and more dithering as well, and it shows a bit in the result (artifacts start to vanish with enough data), but it seems rather removed from OP’s example.
