I have been looking everywhere to find a simple plugin for PS that will down-sample image using a good implementation of Lanczos.
So far I have not found any and I am using a separate application.
I like the results of down-scaling in Genuine Fractal 5. GraphicConverter has Lanczos 3 and is really cheap, not to mention that batching in GraphicConterter is probably the best I've ever seen - very dynamic, very powerful, very easy.
No problem we Windows users are still surviving in the photography world
Photozoom is S-spline which is not very good for down-sampling plus is is $189.
I am not a miser but just for down sampling in PS instead than in a separate application it starts to add-up.
I will look at Genuine Fractal 5 to see if it can be found at a serious discount versus 6 which is $159 or I might just bite the bullet if the results of the demo are spectacular versus the Lanczos implementation I am using now.
Few algorithms can beat a good laczos for detail retention / apparent sharpness when downsampling, and GF5 is not one of them in my opinion (I have tested both of them, ended up with my old LR2/PSCS3/ImageMagick workflow again.)
I've tested X-file2 also, and found that lower in quality than IM too...
A SinC with sufficiently high radius with perfect sigma-matching to the subject - will vary from picture to picture - will do better both visually and mathematically, but really isn't worth the hassle. Catrom is the same, and generally needs pre-averaging with a radius determined by the down-sampling ratio.
I haven't found a good Lanczos to use within PS either, and I spent a while searching.... Please report back if you find anything worthwhile!
It's -5ºC and wet/icy here, so I think I prefer something with a hardtop right now... Looks good for summertime though! Hope you'll enjoy, and drive safe...
The important part of the OP's q though was "usable from within PS".... I agree to the suggested solutions apart from that, I go outside Adobe to resample my pictures myself.
It's -5ºC and wet/icy here, so I think I prefer something with a hardtop right now... Looks good for summertime though! Hope you'll enjoy, and drive safe...
The important part of the OP's q though was "usable from within PS".... I agree to the suggested solutions apart from that, I go outside Adobe to resample my pictures myself.
I am in agreement on all counts
Would not drive too much on snow.
I am also going out of PS to IDimager which allows me to batch Lanczos down-sampling and resizing and then go back to PS for a touch of Smart Sharpen.
I was just looking at staying in PS and an action to do the same.
Here's a repost of what the IM Lanczos3 can do compared to PSCS3's internal "Bicubic sharper".
The original picture is downsampled to an uneven ratio, ~43% or something - I don't remember - and then scaled up 200% with "nearest neighbour" blocks to amplify the differences. No sharpening applied anywhere! I also find it interesting that the Lab BiCubic is so much better than the RGB, inside Adobe... But then we've always known that PS's RGB-handling is pretty crappy (from a purely technical point of view).
A little trick when downsizing. Do a slight gausean blur (you need to experiment; I used 2 in GIMP), downsize, and then re-sharpen. Helps with the aliasing issues (stairstepping).
I've given up hope that Adobe will ever incorporate Lanczos. They may as well call it BicubicShop for cripes sake. As such, I've found no plugins to do the job that I've felt were worth any money. I settled on FastStone Image Viewer, which is a good all around utility with lots of options in the algorith department, though only one implementation of Lanczos, which I believe to be Lanczos 3x3 after comparing it with Picture Window Pro which has a few variations. FastStone is free (donations accepted), gets somewhat regular updates, makes a very good file viewer and has a superb batch processer built in, so it tends to make itself quite handy in many regards.
After many hours of searching and trials that's what I've come to.
lylejk: try again - but this time downsample the 480x480 original to 200 pixels (41.7%), and show your result at 200% scale (with "nearest neighbour" scaling). This is what I did.
Using a Gaussian blur at [1/sampling ratio] radius (following this generalized rule would give 2.4 for this example) is a good way to avoid aliasing effects or actually algorithm deconvolution artefacts when downsampling, but only when the chosen algorithm induces artefacts (as bicubic does)... And you have to manually eyeball the amount of sharpening added after resampling to make the result as good as possible. Even after this amount of work, you end up with artificial detail, as what you're sharpening isn't necessarily there in the original, it might be an artefact created by a bad resampling algorithm...
I'm not arguing that a bicubic can't be made to LOOK good with a certain amount of work, but I for one want to avoid that work. And for my internal peace of mind I want to know that detail I see in the finished, resampled picture is real, not something invented by photoshop - but that might just be me...
Knowing that lanczos preserves more of the original image than bicubic (both mathematically and visually) is reason enough for me to go out of my [PS] way to use it.
Ahh; sorry about that. I really need to read more carefully. lol
I actually used Lanczos sampling. It does a very good job, but this technique still helps here too. Your right about the eyeballing (both blur and sharpen steps). I'm hopeing this guy will develop a better technique for downsampling as he already has for uprezzing; photocomix on GIMPTalk said he was going to tackle it. We shall see, but Lanczos is the king imo.
The endorsement of Lanczos by Joakim can be called ringing, but "ringing" is also the problem. The sinc function, being related to the Fourier transform, can preserve information to remarkable effect: lossless rotation of any angle, for example. The intermediate versions of images so transformed will look pretty bad, unfortunately, with schmutz everywhere. For this reason, and for computational efficiency, sinc is windowed (i.e. sampling localized) using various homebrew techniques (Blackman, Mitchell, Lanczos, etc).
When first writing my raw-processing tool, I thought Lanczos was the cat's meow, and made it my only interpolation function. It became clear that high-contrast points and edges had radiating lobes and concentric rings which were noticeable. Even at lower contrast, or with low-pass filtering, the effect was present. This is the nature of sinc, and not subject to fixing.
The final step to my enlightenment was thinking about edge enhancement. This feature is the superficial appeal of Lanczos: free sharpening. This is the false god! Besides the ringing, Lanczos operates in the vicinity of the sampling interval, regardless of the window size. Sharpening as a photographic (really visual perception) matter is necessarily calibrated to the eye, to its edge-detection foibles, and the viewing geometry.
I find bicubic perfectly acceptable as an interpolation function, even if it's destructive mathematically. The image can then be sharpened - like spicing your soup - to taste.
