Comparing downsizing algorithms

Started 7 months ago | Discussions thread
ForumParentFirstPrevious
Flat view
Digital Nigel Forum Pro • Posts: 12,900
Comparing downsizing algorithms
3

Modern digital cameras produce much higher resolutions than are needed for on-screen display, or even normal printing. So the images need to be downsized using resampling. But which algorithm to use?  I looked at four:

  1. Bicubic would be the normal default algorithm, available in most photo editing packages. It's fast on modern computers, and produces generally acceptable results.
  2. Bicubic Sharper is a newer variant, designed, as the name suggests, to produce sharper downsized images. But they can sometimes have an over-sharpened 'crunchy' look, so it's best used with caution.
  3. Lanczos is a newer algorithm that's slower, but believed to produce slightly better results. It's still very fast on a modern machine.
  4. Gigapixel AI is a proprietary product that claims to use AI machine learning to produce much better results than these mathematical algorithms. In its latest release 5, it offers a new 'Man-made' option, trained on man-made structures, previously its weak point. It's very slow to operate, perhaps taking a hundred times as long as any of the algorithmic options.

I decided to put them to the test with this image, that I took in the Louvre Museum last year:

Processed  from raw using DxO PL3, full-size (after volume deformation correction and cropping); 7211x4810

And now, the same images downsized to 4000x2668:

Bicubic, exported directly from PL3

Bicubic sharper, exported directly from PL3

Lanczos 3, exported from Affinity Photo 1.8

GigaPixel AI 5.0, using the new 'Man-made' mode

Comparing them:

Comparing the results from four downsizing methods

I'm curious what other people think of these results. For example, are any of the more exotic options any better than good old bicubic, which is fast and widely-available?  Do you think the newer Bicubic Sharper look over-sharpened ('crunchy')?

Even if you have Gigapixel for upsizing (where it can do a very good job), is it worth using it for downsizing, given that it's more cumbersome to use, and very, very slow?  In fact, is it any better than the cheaper, much faster alternatives?  I'm certainly not convinced.

In fact, overall, I suspect that the good old Bicubic algorithm actually did the best job here.

 Digital Nigel's gear list:Digital Nigel's gear list
Panasonic FZ1000 Canon PowerShot G7 X Nikon Coolpix P900 Panasonic ZS100 Sony RX10 III +19 more
ForumParentFirstPrevious
Flat view
Post (hide subjects) Posted by
ForumParentFirstPrevious
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow