Luminous Landscape Forum
Raw & Post Processing, Printing => Digital Image Processing => Topic started by: Benny Profane on February 19, 2023, 04:27:06 pm
-
Does anybody use super resolution in LR as much as I do (a lot, first thing)? I've been noticing a weird pattern showing up in areas lately. Maybe they were always there, but, who knows.
Both samples come from the raw landscape and sky image. One, super resolution, one not. Very very zoomed in, it's a chunk of the mountainside just a little lower left of the center flat peak, so, yeah, maybe no whoop to the naked eye, and I can't see it in prints, but, still upsetting.
This does not happen on all images.
-
Upload a DNG with existing edits/profile to something like Dropbox so it can be examined.
-
Does anybody use super resolution in LR as much as I do (a lot, first thing)? I've been noticing a weird pattern showing up in areas lately. Maybe they were always there, but, who knows.
Both samples come from the raw landscape and sky image. One, super resolution, one not. Very very zoomed in, it's a chunk of the mountainside just a little lower left of the center flat peak, so, yeah, maybe no whoop to the naked eye, and I can't see it in prints, but, still upsetting.
This does not happen on all images.
I would avoid Super Resolution and try Bicubic Smoother in PS, Topaz Gigapixel or Raw Therapee's Lanczous algorithm to upres your images. SR isn't perfected yet. I'm going by a processing guru who has tested pretty much every upres method.
-
Does anybody use super resolution in LR as much as I do (a lot, first thing)? I've been noticing a weird pattern showing up in areas lately. . . . This does not happen on all images.
"Artificially intelligent" techniques for increasing resolution based on neural network processing aren't actually enlarging the target image the way a pixel-interpolation approach would; they are generating a new, derivative, image based on visual fragments found in regions of the source image that the neural network is able to recognize and reproduce at a larger size. There will always be artifacts in the target if you examine the source and target images at a sufficiently small level of detail.
Sometimes one neural network will produce a cleaner result with a specific image than another—due, I suspect, to the type and number of images that were fed to the programs during their respective training phases as well as the way the different software products were coded. I've occasionally tried Lightroom's Super Resolution and Topaz Gigapixel on the same source image; sometimes one of the programs does a demonstrably better job, sometimes the other. At the same amount of "enlargement," the differences typically are negligible except on very close inspection. (Note that, unlike Super Resolution, Gigapixel isn't limited to 2X linear resolution increases.)
All these computational imaging techniques tend to improve over time thanks to continued training with additional images. I would be surprised if Adobe wasn't continuing to train Super Resolution, Topaz with Gigapixel, and other providers with their "intelligent" imaging tools.
Very very zoomed in, it's a chunk of the mountainside just a little lower left of the center flat peak, so, yeah, maybe no whoop to the naked eye, and I can't see it in prints, but, still upsetting.
If you can only detect the artifacts when the image is "very very zoomed in," and you can't see them in prints, my recommendation would be not to worry about the phenomenon.
-
Yeah, Adobe is always working on stuff, I guess I caught a glitch in what will be an early version of SR. As long as there is no significant cropping, basically, no whoop, I guess.