Why the sky turns into stripes

Two hundred and fifty-six steps, and where they run out

An 8-bit channel has 256 possible values. In a photograph of a market, a street, a face, that is plenty: the tones jump around from pixel to pixel anyway and you will never see the steps.

A smooth gradient is different. Take a clear sky that runs from value 150 at the horizon to value 180 at the top of the frame across 1,200 pixels of image height. That is 30 available levels spread over 1,200 pixels, so each level occupies a band 40 pixels tall. The human eye can reliably detect a one per cent difference in luminance in a smooth area, and one level out of 256 near mid-grey is about half a per cent: close enough that under the right conditions you see the edges. You get banding: visible stripes where the software had no value in between.

Now stretch that sky. A Curves move that expands those 30 levels to fill 90 levels does not create 90 distinct values. It maps the 30 you had onto 90 slots, leaving 60 of them empty. The bands get three times taller and three times more contrasty. Open the histogram afterwards and you will see it: a comb, with white vertical gaps where no pixels remain.

Where it bites in practice

• Skies, especially blue ones, because the blue channel is the noisiest and gets the heaviest smoothing in JPEG.
• Studio backdrops and seamless white, which are pure gradient by design.
• Dark scenes. Because of gamma encoding, the shadow end of an 8-bit file has fewer code values per unit of actual light than the highlights do, so lifting shadows bands faster than pulling down highlights.
• Flat colour behind a logo, where a subtle gradient in a brand background turns into terraces on a printed banner.
• Anything shown on a big screen, because bands are a spatial artefact and a projector makes each one a hand’s width across.

What 16-bit actually buys you

Switch to 16 bits per channel (Image > Mode > 16 Bits/Channel in Photoshop and Photopea, Image > Precision in GIMP, and Krita offers 16-bit integer and 16- and 32-bit float) and each channel gets 65,536 levels instead of 256. That is 256 times the room. Any stretch you apply still leaves gaps, but the gaps are far smaller than one 8-bit step, so when you convert back down at the end the result is smooth.

Photoshop’s “16-bit” is actually 15 bits plus one, giving 32,769 levels rather than 65,536. It is a historic implementation choice, it makes no practical difference, and it confuses people who check the maths. GIMP and Krita use the full range.

The costs are real and worth naming:

• The file doubles. A 24-megapixel 16-bit TIFF is around 140 MB. A 30-layer document at that depth will exhaust a laptop with 8 GB of memory.
• Some filters and plugins are 8-bit only. Older GIMP plugins and a few Photoshop filters grey out in 16-bit.
• Photopea’s 16-bit support is partial. It will open and convert, but several adjustments internally round to 8 bits. For heavy tonal work on a gradient, this is one of the places where the free browser tool falls short and GIMP or Krita is the better free choice.

Dither: noise as a repair

There is a second defence, and it looks like the opposite of quality. Adding a small amount of random noise breaks the boundary between two bands into a scatter of pixels from each, and your eye averages them back into a smooth transition. This is dithering, and it is why film grain hides banding that a clean digital gradient shows.

Practically: Photoshop’s Gradient tool has a Dither checkbox: leave it on, always. GIMP’s gradient tool has the same option. When you flatten a 16-bit file to 8-bit for delivery and still see a band in the sky, add a noise layer at 1 to 2 per cent, set to Overlay, masked to the sky. One to two per cent is enough. It will survive JPEG compression better than the band will.

The limit nobody mentions

Converting an 8-bit file to 16-bit does not recover anything. If you open a JPEG and switch it to 16 bits per channel, the file now has 65,536 available levels and still contains only 256 distinct ones. The histogram is the same comb it always was, just drawn on a wider axis.

What the conversion does is stop you adding more damage: further edits now round into the finer grid instead of piling up on the coarse one. That is worth doing before a big tonal move on a JPEG, and it is not the same as having shot 14-bit raw. The only way to have the levels is to have captured them.

One more honest note: the final delivered file is almost always 8-bit, because JPEG, WebP and the web are 8-bit. Working in 16 is about protecting the process, not the output.