Exactly as Chris said, the pre 3.6 version lead to very bad results only when zooming in and people complained repeatedly about this. (When exporting everything was fine).
Unfortunately the algo works well only if large parts of the raw image data are used and because of technical reasons hidden in dt’s processing pipeline that couldn’t be fixed yet (hopefully i’ll find a solution for 3.8) At least you get a warning that CA is bypassed atm.