Friday 18 April 2008

It's bigger but is it better? A follow-up

Rock snake, Northumberland, March 2008
crops below are from near the centre


It's now over a month since Ctein posted a series at T.O.P. on enlargements (part 1, part 2, part 3) an I promised I'd follow up on my idea of using mixed methods.

To summarise what's gone before: Ctein looked at various enlargement techniques, especially comparing Photoshop Bicubic (BC) with Genuine Fractals (GF). The overall conclusion - BC retains detail much better for bigger enlargements but returns softer results. GF is aggressive about sharpening edges but can lead to loss of detail and polygonization. Up to an enlargement factor of 2, GF was tending to win.

I ran my own quick comparisons between the two and got the same results - GF particularly likes turning fine, random detail into lots of little triangles that can be quite irritating, but bicubic requires a lot of extra sharpening.

So, I thought, could I get the best of both worlds? Could I get the fine detail of BC with the sharp results of GF? My initial hypothesis was that GF could only "see" details once they got to a certain size, otherwise it assumed it was noise or something. If I could initially make those small details large enough, the GF would see them and work the magic.

What did I do - I ran a lot of tests and a bunch of prints. I was looking at several variables here: relative degree of enlargement between the 2, absolute degree of enlargement and overall combined enlargement. That leads to a large solution space.

In the first instance I a whole lot of blind enlargements for totals of 400% and 800%. I ran relative enlargements of 1-0 (one method or the other), 2-1 (one method with a % enlargement twice that of the other) and 1-1 (equal enlargement %). Then, from those findings, I played around in the areas where the best results were happening. I repeated over several images, covering fine detail, geometric patterns, in & out of focus regions. Always BC first (see aside at the end for more on this).

This is the summary of results (without boring you with a lot of intermediate stuff). I've tried to keep my workflwo as consistent with Ctein's as possible but generally I worked with crops for prints (I'm limited by printer size), only used the straight BC (very little difference for this purpose between straight, softer & sharper), and worked with GF 4 (not v5 which again, for these purposes, is not a major factor).

At the extremes BC alone is soft, GF is sharp but blocky. In between there is a continuum of results. If too little BC is used to start then the polygonization of GF starts to rear its ugly head. If too much then the results become too soft. The continuum is not a straight line, however. As i ran more and more enlargements and prints, I found that starting with a BC enlargement in the 150% to 200% range yields the best results. Below this, there are noticeable GF artefacts, above this results start getting soft. The softness starts to creep into the out of focus areas first - with a BC enlargement of 250% to start, any slightly off focus area is quickly (and unattractively) rendered soft. Go further and everything goes soft. GF cannot retrieve softness induced by BC.

Crop from above image of a detailed, in focus area. Total Enlargement 400%
L to R: BC 150%, 200% and 250% (thus GF, 250%, 200%, 160% respectively)
Click to get full-size 100% view.
Differences on screen are small, trust my write-up, not these crops.


This is almost independent of total enlargement. Only when the GF enlargement is great (more than 400%) can I notice a difference between the BC 150% and 200% results - on screen or in print. More BC helps here. Frankly, I wouldn't be enlarging digital work more that 800% total anyway - that would be 48" prints and more. These tests are looking at prints that would cover a wall from a distance of a few inches.

Another crop from the same image.
This is a slightly out of focus area (by a few millimetres). Same enlargements as the other crops.
The 250% BC image on the right is starting to blur noticeably.


My conclusion (and now my default workflow): start with a BC enlargement of 160% (just over the 150% minimum threshold) and use GF for the rest to the desired size. want to go more that 800% overall (lunatic) - start with BC 200%.

An aside - starting with GF. A tried, briefly, the enlargements the other way, too - first GF ten BC. Don't bother. The result is you get the worst of both worlds. First GF strips the detail and turns it all into little triangles which BC enlarges and softens. It's like some nightmare special filter effect.

3 comments:

  1. Very cool! I want to give this a try.

    Which bicubic were you using? Bicubic, bicubic smoother, bicubic sharper?

    ReplyDelete
  2. Joe: just straight bicubic (I think I mentioned that).

    ReplyDelete
  3. Joe, if you experiment with BiSmoother instead of BC, please post results here or on your website? Thx!

    ReplyDelete

I like comments, especially constructive ones.
Comments get emailed directly to me before publishing , so if you want to get in touch drop a comment.
All comments moderated by me before being published, keeps the spam at bay.