Bifurcator wrote:
A sharpening contest?
Perhaps not a contest, since there is room for personal taste in how an image should be presented, but at least a consistent side-by-side comparison from which one can draw their own conclusions (about who wins ).
The solution I take for overly aggressive sharpening tools (and they all are in my opinion!) is at least in the case of Photoshop, to do a fade operation after. This allows you to see and compare before and after as well as adjust the affects of the tool to suit each image. Here's my script as it is today...
Instead of converting to Lab and back and using Fade, I create a new layer that combines all layers below (a copy of the entire image), set the blending mode to Luminosity and set the opacity % slider. Works the same way, but I can change the "fade" later (as long as I don't flatten the image).
mpmendenhall wrote:
This might make an informative comparison:
If someone has a nice, detailed, minimally-PP'd full-size image that they can post a link to here, and then everyone else can apply their favorite downsizing/PP method on the same image (or set of a few images) to the same target size (e.g. within 800x800) and post the results in this thread. Let's see how these different workflows look! I'll try to dig up some sample image tomorrow morning if no one else has posted one to work on by then.
Not sure if it meets all requirements, but here is very detailed 21MP raw from the new ZE 35/1.4:
Mirek Elsner wrote:
Not sure if it meets all requirements, but here is very detailed 21MP raw from the new ZE 35/1.4:
elsners.org/misc/35-1-4.zip
Thanks, Mirek! That looks like a great sample for this test.
Target size is 800 pixel width.
To just compare re-scaling algorithms rather than other PP, I've refrained from additional PP adjustments to the image; in a "real life" case, I would pull back the highlights to show more of the cloud detail.
Next, ImageMagic results, scaled from full-size jpg output from Aperture. More aggressive detail sharpness without "halo" artifacts, since the sharpness is produced by under-sampling (like a camera with no AA filter) rather than edge-contrast increases: http://praetoriusphoto.images.s3.amazonaws.com/fmforums/20110503_downsize_im.jpg
Look closely at the shingle roof of the house right of center; Aperture has made a mess of it, while the ImageMagick resize shows the shingle pattern. Also, the Aperture output is ~2x as large filesize as the ImageMagick (both adjustable parameters), but doesn't have anything to show for it. The Aperture version also shows some slight edge halos, from whatever sharpening is built into it; I have seen some really bad edge artifacts from Aperture exports.
Hmm, that's weird. You said you did not do any additional PP? That image is waaay overexposed compared to what you get when you develop the image with default parameters in Lightroom. Here's what I get when I develop the raw file without touching the settings, and running a standard resize script:
Anyway, there's one thing in the discussion that has been overlooked so far - the fact that the camera that produced the image and the monitor you look at the image on are as important as the sharpening/resize algorithm. For instance if I run my 5DII script on my M9 files they really get damaged by the script (21 Mpixels vs 18 Mpixels & no AA filter).
The sad truth is also that the monitor dotpitch plays a huge role. Something that looks bad on my 24" 1920x1200 workstation monitors may look good on my 17" 1920x1200 laptop monitor and vice versa. So the target device matters.
denoir wrote:
Hmm, that's weird. You said you did not do any additional PP? That image is waaay overexposed compared to what you get when you develop the image with default parameters in Lightroom. Here's what I get when I develop the raw file without touching the settings, and running a standard resize script:
The file has a non standard profile named "5d2 large cc wnd" assigned and the standard values of the file have been changed if opened in LR.
So you are absolutely right, Luka, the file is far from no additional PP.
You have to reset in LR and you'll see the overexposed version posted above.
denoir wrote:
The sad truth is also that the monitor dotpitch plays a huge role. Something that looks bad on my 24" 1920x1200 workstation monitors may look good on my 17" 1920x1200 laptop monitor and vice versa. So the target device matters.
I agree with this. The larger the pixel pitch, the sooner you'll be noticing oversharpening artifacts, I believe. I've found 1920x1200 at 24" too low a resolution, especially when I first started using it, being used to 2560x1600 on an ACD 30" (of course that wasn't a voluntary switch ).
I think everyone who uses step sharpening scripts created by others should adapt it for use of their own camera, in case it's not the same as the one it was created for. Especially if it's a very different type of camera (M9 vs 5Dmk2).
Mirek Elsner wrote:
Instead of converting to Lab and back and using Fade, I create a new layer that combines all layers below (a copy of the entire image), set the blending mode to Luminosity and set the opacity % slider. Works the same way, but I can change the "fade" later (as long as I don't flatten the image).
That's a good idea. But it's about 3 or 4 times slower. Doing any operation on a layered PS file - especially a pixel operation - is gawd awful slow. I can barely put up with the NR step here. This would be better (maybe) if it were just one or two final images rather than my usual 20 to 50 image batch.
Morfeus wrote:
The file has a non standard profile named "5d2 large cc wnd" assigned and the standard values of the file have been changed if opened in LR.
So you are absolutely right, Luka, the file is far from no additional PP.
You have to reset in LR and you'll see the overexposed version posted above.
Heinz
I guess if we want a comparison of resizing algorithms, and eliminate other variables, it would be better to all start from the same .jpg instead of a RAW file.
And here is a more strongly sharpened version, by adding moderately large radius unsharp mask sharpening before the rescaling: convert 20110504_scaleme.jpg -quality 98% -interlace line -filter gaussian -define filter:support=0.6 -unsharp 5.0x3.0+0.8+0.00 -resize 800x800 20110504_scaled_b.jpg http://praetoriusphoto.images.s3.amazonaws.com/fmforums/20110504_scaled_b.jpg
Hmm, that's weird. You said you did not do any additional PP?
That's actually my fault. The image is exposed to the right and I should have said that. In some converters in may be necessary to lower the exposure by 1EV if you don't use my LR presets.
Finally, edging into "overbaked" sharpening (but some people like it that way): convert 20110504_scaleme.jpg -quality 98% -interlace line -resize 2400x2400 -filter gaussian -define filter:support=0.5 -unsharp 3.0x2.0+0.8+0.01 -resize 800x800 20110504_scaled_c.jpg http://praetoriusphoto.images.s3.amazonaws.com/fmforums/20110504_scaled_c.jpg
This process pre-scales the image down to 2400 width, before applying USM and finishing the scaling to 800. It takes <4s to run on my four-year-old MBP, so it's reasonably fast.
mpmendenhall wrote:
I guess if we want a comparison of resizing algorithms, and eliminate other variables, it would be better to all start from the same .jpg instead of a RAW file
not a good idea IMHO.
Most of us would start with a pre-sharpen in LR, Aperture or C1 and then go to PS and do the resize and output sharpen step.
I think we should use the raw, reset it and then start all over.
I put my pictures on my web and allow people viewing in small size or full screen. So some of the sharpening is done on the server dynamically. So I don't know the final size and I can't use any sophisticated PS algorithm for that.
My first sharpening example is from that sharpening for my website. I just convert in LR and export with size constrained to something like 1900x1200 with sharpening set to Standard. I import the picture to my gallery and it is then dynamically resized for viewing by the server using Lanczos.
Regular size is 840, you can click the rightmost icon on the black bar below the picture to see full size.
Most of us would start with a pre-sharpen in LR, Aperture or C1 and then go to PS and do the resize and output sharpen step.
I think we should use the raw, reset it and then start all over.
Maybe, but then I think we would end up more comparing the differences in RAW rendering color/saturation/blackpoint/highlights defaults between converters (which would be an interesting different test on its own). The full-size .jpg version should contain most of the information from the RAW, aside from extra latitude for levels adjustments; you can probably do adequate pre-sharpening preparation from there, especially for the large downsizing ratio where most of the difference will depend on the downsizing algorithm rather than pixel-level details of the 24 megapixel image.
Resize to 800px using bicubic
Smart sharpen radius .25, strength ~300 on a separate layer
Blending mode set to Luminosity and Opacity to 2/3 (65% or so)
And the same again, but with Opacity at 40%. My goal with this one was to get close to Luka's with simple processing. Luka's show more local contrast. I will see if that can be fixed with USM with large radius or something later today...
).