Downsampling question.

[Steve Moore-Vale]

When I resize an image down from its full resolution to one for sharing online, I use 2048 pixels longest edge, I always tick the downsample box and choose bicubic sharper. When I do this the image preview (showing the smaller version) looks noticeably sharper than when I untick the box. Is there some sort of sharpening going on just by choosing Bicubic Sharper or is it always worth sharpening the resized image with something like smart sharpen after resizing?

[D Fosse · Accepted solution]

Yes. Bicubic Sharper does indeed have built-in sharpening - and a little too much for my taste. I prefer Bicubic Smoother, which hardly sharpens at all, and then run the ACR filter for final sharpening at output size.

BTW Bicubic Automatic is just Sharper down, and Smoother up.

[rayek.elfin]

Yes, I agree with D Fosse: Photoshop's Bicubic Sharper doesn't yield good results due to too much built-in sharpening.

In my experience, and testing done by others, Catmul-Rom gives some of the best downsampling results of all downsampling algorithms, but you guessed it: Photoshop doesn't include it. Which is a real shame. Which is also the reason why I avoid Photoshop for any downsampling work.

[Trevor.Dennis]

rayek.elfin  wrote

In my experience, and testing done by others, Catmull-Rom gives some of the best downsampling results of all downsampling algorithms, but you guessed it: Photoshop doesn't include it. Which is a real shame. Which is also the reason why I avoid Photoshop for any downsampling work.

That's a new one on me, and it has apparently been around since before Photoshop was born.  Google tells me that Imagemagik makes use of it, but it is not altogether clear.  If you know of a Windows version that makes use of it, I suspect I am not the only person who'd be interested.

[rayek.elfin]

Here is an interesting article on this topic:

PixInsight Reference Documentation | Interpolation Algorithms in PixInsight

I did (a lot!) of my own testing, and Catmull-Rom works particularly well with downsampling sharp-edged details and vector-based artwork. Mitchell-Netravali also works quite well for these tasks, and delivers a slightly smoother result. The end result may be improved further with additional sharpening - the point being that the source down-sampled image is far more controllable at that smaller size (Photoshop instead decides for the user to over-sharpen when using the Bicubic-sharper setting, and bicubic is too fuzzy for downsampling).

Applications which support Catmull-Rom and Mitchel-Netravali downsampling are:

• Color Quantizer
• PhotoLine
• ImageMagick (already mentioned)
• PixInsight (see above)
• Krita (only Mitchell-Netravali)

Color Quantizer offers a ridiculous number of resampling algorithms, both with in and out sharpening. I tend to use either CQ or PhotoLine for my downsampling needs. It's an amazing free little PNG optimizer.

Color quantizer

PhotoLine includes a nice option to control the sampling algorithm on a per-layer(!) basis. Very handy since all layer transformations are by default non-destructive.

[Trevor.Dennis]

As Dag has said, Bicubic Sharper is the work of the devil, and best avoided if you have a have decent original image.  Ì set Preferences to Bilinear, and sharpen after downsizing _if_ required.

[D Fosse]

In most cases this is a solution in search of a problem. There are many challenges during a normal workday, but this isn't one of them.

The critical part is to control the final-size sharpening, meaning the resampling itself should do as little as possible of it. Bicubic Smoother is fine.

There is one possible exception: Thin lines - telephone wires, suspension bridges, that sort of thing. That too usually works fine, but in some rare cases a little special care is needed. Sometimes a two-step downsampling improves it a bit.

Still, going from 7K pixels to 1K or so, something has to go.

[Steve Moore-Vale]

Thanks for all your replies. Very much appreciated and very helpful. I will avoid using Bicubic Sharper in future then as it seems that it is not the best option to use. A friend of mine one told me he resizes for sharing online and then creates a new layer and adds a smart sharpen filter to it but I find that too agressive as well. I do mostly bird photos so a lot of fine edges in the birds feathers. Plus aviation too where there’s fine detail on the bodywork but also lots of plain panels with no detail.

[Stephen Marsh]

There are various ways to sharpen with more control (Unsharp Mask, Smart Sharpen, Adobe Camera Raw) than using one size fits all methods… When you sharpen on a separate layer/file, the result can be merged back over the original image in many ways:

* Reduced opacity

* Blend modes

* Layer option blend if sliders (link to my old website)

* Tonal based layer masks (standard/inverted/bell curve)

* Edge isolated layer masks (link to my old website)

* Hand painted layer masks

P.S.: Some of you may like to try the following:

C3C Image Size – Telegraph

[D Fosse]

Yes, the separate layer sharpening is useful for all the reasons Stephen mentioned.

Personally I found the ACR filter made most of that redundant. All of those blending effects can be achieved in one go in ACR, and in many cases even better.

The main thing is to avoid halos and other artifacts. Starting out with Detail=0 in ACR does that, it tightens up edges only, at the given radius. Then add Detail carefully to taste, and masking to protect flat areas.

But sharpening is a matter of personal preferences. Which is exactly why the Bicubic Sharper algorithm should be avoided.