Jonathan Turner asks:
(questions that requiring pretty complicated answers, bear with me for this
long winded reply, or go to #1 for the short version below ;)
The point about viewing distance is a good point, which I
hadn't really considered.
Viewing distance has a heck of a lot to do with everything in photography,
up to and including lens choice when making the image. Wide angle images
which look all distorted can look perfectly normal if enlarged to wall sized
and you force the viewer to stand 2 feet from the image - similarly a 1000mm
telephoto shot which we normally describe as 'compressed' will also look
totally normal if printed appropriately small OR you force the viewer to
stand a long way back (viewing angle the same).
Also, increasing the image size in 10%
increments is something I'd heard about
before but forgotten. I tried it
actually and I think it was was slightly better.
it should be slightly better, but not much - and if you experiment and start
with a clolour swatch you will find it eventually creates patterns - highly
undesirable. I illustrated that in the webpage I liked but anyone can test
it themselves, upsize in degrees until a blindingly obvious pattern is
created. resizing is bot a good way to make things bigger, resampling is
much better. As long as the algorithm is appropriate (b-spline, s-spline,
triangle, hermite, Mitchell, Lanczoz, Bell, etc etc)
similarly downsizing (as opposed to downsampling) often creates patterns -
more often moire .. from unintelligent bit removal.
What you appear to need is some serious upsizing. If this were being sent
straight to print without mucxh manipulation, I would suggest you'd be
leaving the upsizing to a RIP, something that's been discussed often in the
past on this list. However the process of incorporating image elements with
differing levels of detail demand a more complicated pathway.
The real issue though is the disparity between the background image and
the composite elements ...
see way down for more on this (#1)
as such I want to try and make the background image the best quality I
can to start with. So, I went back to the original RAW files, and
remembered that you can open them at different sizes/resolutions...at
the bottom of the RAW window in Photoshop you can specify the
dimensions/resolution and Bit depth you want to work with.
Photoshop is weird. I really wish they would stop mixing ancient printing
terminology like dots per square inch in (irrelevant) with computer tech -
pixels. Your image has a native resolution and if shot RAW it'll be the
full sensor size. I will asume it's a 12Mp camera and that'd mean you have
something like a 3k x 4k pixel image. dpi is irrelevant BUT photoshop
wants to know what sort of silk screen you'll be squeezing your oil based
ink through and thus demands to the know the dot pitch frequency (bah) Tell
it 300, then check that your image is still only 3k x 4k pixels in size when
it opens the image.
Dimensions and resolution I understand,
I apologize if I come across as suggesting you may not - I've found a lot of
confusion starts and never gets past Photoshop forcing weird conflicting
concepts of 'resolution' on people so I try to define the problem so we're
on the same page ;)
though Bit depth is something
I'm unsure about...I have the choice of 8 or 16 bits,
simple answer is 'work in 16 bit'.
Bit depth is like the number of steps between absolutely no colour and total
colour. black (no clolour in transmitted) to white (all colour in
transmitted) - a bit of basics (sorry!)
if there are 4 steps you have black, dark grey, grey, light grey, white.
traditionally we have 1 bit (1 to the power of 2 or 1^2 = 2 colours.. ie,
black/white)
2 bit (2^2 = 4 colours)
4 bit ( 2^4 =16 colours)
8 bit (2^8 = 256 colours) - this is a point where we
16 and above bit gets a bit weird, with some padding bits needed for
computational purposes being added in some cases (see here
http://en.wikipedia.org/wiki/Color_depth#High_color_.2815.2F16-bit.29 ) but
basically for editing, more is better - if your exposure is damned good to
start with and you're not going to stretch out the tonal range, 8 bit is
fine (no different from shooting E6) - and although their are now printers
that apprently work with 16 bit, I've not seen the output and would suggest
that while they might 'handle' 16 bit, they may (or may not, I truly have no
idea) simply be data converting and still printing using an 8 bit process.
What this is all about is how good our eyes are at perceiving the
transitions of colours - remember ever seeing pictures where the sky broke
up into distinct bands of colour? Often that was through over-editing or
more frequently, using a low bit depth image. If you only have say 16
colours to represent the sky, it'll break up into bands of colour to
represent the closest colour available to reproduce the actual sky colour.
You don't see this much any more..
However it comes down to, the human eye is hard pressed to see discern
1/256th of a shade of colour difference in an image. ie, 8 bit (256 shades
of individual colours) arranged in a line looks like a smooth transition
with no idividual patch standing out as being darker or lighter. Actually
it is hard pressed to see 50 gradation swatches in mid tones, but eh.. 6 bit
colour isn't really that mainstream (64 steps, and we actually can see
gradations in tarker tones better than lighter tones)
When it comes to image manipulation however, the more the merrier if you
need to wriggles those tones about. stretching out the tonal range is a lot
less likely to show banding if you have 65000 gradation levels to expand out
compared to 256. I could write 100 pages of explanation but you'll see it
yourself if you experiment..
and I'm presuming
if I choose 16 bits it will be better quality,
nope, just more of a gradation of tones between black and white, only of use
if you fiddle the levels enough to find banding occurring in 8 bit, and not
in 16 bit.
I've tried using it and it
certainly increases the files size...so much so that when I have a final
composite psd the file size exceeds 2gb..
if I choose a higher resolution setting than
300PPI will that actually increase image quality?
I think you're fighting the Photoshop File Size Mangling Feature - it's been
an age since I fought this beast but the essence is that you will probably
find you have inadvertently resized the image and added data. Let me see
if I can remember howq this fight goes - I saw students doing this often..
It kinda went like: Open a camera image of say 3000x4000 pixels (which the
camera may have tagged 72 dpi for some insane reason), change the dpi to 300
(if you don't think this through, your may find your image is now 12,500 x
16,667 pixels - 17x bigger .. {or (300ppi /72ppi) squared bigger). Add a
layer to that and it gets worse, all of a sudden you're losing this battle.
300 pixels per inch (ppi) is an acceptable viewing resolution at 'normal
viewing distance' (1/2 arms reach) for print - a compromise, a general
quality standard. heck, most monitors are only 90-140-ish) (or 72 ppi if
you talk of the early macs) At 300 print pixels per inch most people don't
see blocks - for some even 100 is OK, but truth is, cut out or blank one
line of pixels in an image (or add one line of white pixels) and even at
1200 pixel resolution you'll see it. heck, drop a line from a 4800 dot per
inch (dpi) printer with a blocked jet and you'll see it. but that's a line
and we recognise repetition (and lines) better than teeny individual dots.
Any good tips for improving overall base image quality much appreciated.
Just to re-cap; the images are composites with a final output size of
around 1600mm x 1064mm.
you're looking at around 64 x 42 inch image. I'll use inches since we're
using dpi (ppi) even though I'm a metric person myself.. so the ideal image
size that you will be sending to print in pixels is 300(ppi) x 64 inch times
300 (ppi) x 42 inch or 19,200 pixel x 12,600 pixels, or 241 Mp. Something I
think may exceed your camera's original output ;)
(#1 - the process)
So basically you are starting off this process by upsizing your image to
final print size, preferably using an appropriate upsampling algorithm -
save this file. it's big (241 Mp), then forget it for now.
Next we need to discuss 'resolution' in terms of resolving power and not how
many image elements exist in a given dimension (which is how resolution is
usually discussed in digital photography). and this is the dilema you're
facing - you have nice sharp face pictures you're trying to insert on a
background that is slightly less sharp and you want it to look natural -
Resolution is the ability of the system to resolve detail. Like say you
shoot an image and oh yes, the back looks like a nice white wall in image A,
but in image B the camera was able to 'resolve' the background and you
notice it's not a plain white wall, it's actually tiled and the individual
tiles can be discerned. How it was resolved may be a better lens, less
shake, a different focal point, a different aperture - whatever.. it may
even have been a different media.. say a finer grained film with higher
capacity for resolving (resolution) -
Putting these pictures together is going to be tricky as you know, much as
mixing 70mm film and 8mm film would be tricky.. The way to do it is you've
got to work to your limiting resolution, ie the background picture. Now you
can either do it fiddling about and just seeing what looks right, or you can
do it mathematically, or a combination of the two (have a rough mathematical
idea and fiddle it from there until, it looks right.. this will probably
work best :)
Basically, it's back to the viewing angles!
this will get complicated because of my uncertainty about the original image
sizes .. but basically you have a grainy low res background and sharp, high
res bods to overlay.. so you need to lower the res on the foreground kids to
match the best quality you can get out of the background image!
And i'll warn from seeing many many bad examples of this, mixing camera lens
focal lengths can make this a nightmarish task due to the perspective
conflicts (see those many actress shots where the photographer used a slight
wide and stood close giving her a giant head atop a diminishing body and
teeny feet -try putting that on a 'normal' background at your peril!)
Your limit of resolution is your original background image, upsized to 19k x
12.6 pixels. Although this is your final image, you really need to work
with the original image size and ratios pertaining to that. I'll assume all
images were taken at 3000x4000 (12 Mp)
i'll try doing this with ratios and simple arithmetic ..
Your background let's say is a wall roughly 30 feet long shot from say 50
feet away - I'll assume a camera with a 'normal' focal length and talk no
more about the lenses involved. You plan on arranging your 6 (more? less?)
kids in front of this. you have a 30 foot long wall made up of 4000 dots,
or 133 dots per foot.
how big is a kids face? probably 6 inches across maybe? so you only want
that kids face to be 1/2 x 133 dots wide - say 65 dots (pixels) wide. (ie,
you should be able to fit 61 faces across your 4000 pixel wide, 30 foot
wall. I'm guessing you maybe shot these kids at the same resolution but at
maybe 1/4 of the frame width in portrait mode - so they might be 1/4 of 3000
pixels wide, say 750 pixels wide. So you can see you need to downsample
these face images in this example quite a lot! 750 / 65 = 11.5.. Which
means you want to downsample your original 3000x4000 pixel images to (wait
for it) 260 x 347 pixels ! - Now I know that sounds horrendous, but this
would match the 'resolution' of both images.
Having determined this I would procede to select the head /body/whatever of
each individual kid from the original image at full resolution, cut the non
essential image data (the non-kid part) then resample the image down 11 fold
rather than cutting out tiny images from the small image. These can then be
dropped as layers over your background image (they should be the right size
now if you calculated correctly) and all told even though it'll look a
little fuzzier than you hoped, the images should match both in size and
apparent 'graininess' (resolving power is matched)
hopefully I didnt make a dogs breakfast of the calculations but my inclusion
of the methodology should help guide you if I have
-karl
