With darktable 4.7.0+693~g2a57b25fe3. I resized the window (the larger the difference between export size and view size, the larger the processing difference can be). I then took a screenshot, turned on the new ‘HQ processing’ option for the preview, and took the screenshot again (the difference is when scaling is applied to create the preview: as soon as possible, or only at the end). Sure enough:
Do you think there is any chance that this happens as some sort of by product of the scaling in that while its not direct averaging of pixels it is combining them for a downscaling…there was this post a while back about how DT doesn’t calculate the correct value of an image with a checker board of red and green squares… DT would come up with a simple average value of 128 or something when it should actually be 180. The same image seemed to be handled properly in GIMP. I think there may even be an outstanding use marked to be worked on around this… in any case add in the extreme noise and it just makes me wonder if something around this sort of math processing might contribute to seeing these sorts of differences
I don’t know. I suspect that the modules that apply non-pixel-wise operations, such as blurs, don’t work 100% correct, even if they scale down the radius to try and match the downscaling of the image used to generate the preview. I did not have enough time to investigate, but I suspect that the denoise (profiled) module with its large search radius had something to do with this. You could try disabling it and see if the difference in the image rendering (when zoomed out) occurs to the same extent or not.
Well I’ll be damned, when I turned that 2nd denoise module off, the export looks just like the 100% zoom in darktable, at least to my untrained eyes. I would like to shake your hand, this feels amazing.
I had that module in there because I felt like after the first denoise pass, the water became kinda blotchy with reddish and greenish patches, and I felt that 2nd denoise cleaned those blotches up quite nicely. The downside was the waxy whale, but I’ve been staring at this dang whale for [too long] and I had finally decided I preferred a too-smooth whale to a blotchy whale. I figured I would take the runtime hit and accept it as penance for my bad-photography-sins. Silly me, no sacrifice will atone for 6400ISO.
So it sounds like, if I read correctly, that there’s a risk with large non-pixel-wise operations of creating differences between the display and export images. Is there someone, or somewhere, this information would be valuable?
Thank you so so so much for taking the time to mess around with this picture and help me out, and to the rest of you guys who also played with it and chimed in. I’m so stoked.
I have taken to using two instances for denoise. The first is the chroma preset for profiled denoise. I find adjusting the preserve shadows slider can really improve the result here. Then you can be left with the luma noise… I know it can be targeted in profiled denoise but I like the result of the astrodenoise module for that. I find I can get a good results esp if the noise is the sort of salt and pepper type. For extra color noise the surface blur module which was formerly named bilateral denoise is also good. THe contrast eq can also help in some tricky situations… I’m glad you found the culprit…
