Writing · Comparison

Floyd-Steinberg vs Atkinson dithering on color e-paper

For color e-paper, Floyd-Steinberg is the more accurate dither and Atkinson is the cleaner-looking one. Floyd-Steinberg passes on all of each pixel's color error, so average tone stays true. Atkinson drops a quarter of it, which clips light grays to white and near-blacks to black and gives images more contrast. In our test on the six-color Spectra 6 palette, a flat gray of 230 stayed 230 under Floyd-Steinberg and became pure white under Atkinson. Use a full-error kernel when tone matters, Atkinson for illustration and clean backgrounds, and no dither for text.

Published 2026-11-21. Written for teams preparing content or building image pipelines for Spectra 6 signs and frames. The test below is one we ran for this article on our panel palette; the method and its limits are stated with it.

Why e-paper depends on dithering

A Spectra 6 pixel can only be black, white, red, yellow, blue or green. Every other color is an illusion built from a pattern of those six, viewed from far enough away that the eye blends them. The choice of pattern therefore changes the image more than it would on a screen with millions of colors. Our Spectra 6 image preparation guide covers the full workflow; this article looks closely at the two kernels people most often compare.

Error diffusion: A dithering method that rounds each pixel to the nearest available color, then spreads the difference onto pixels not yet processed, so nearby pixels compensate for it.

How the two kernels differ

 Floyd-SteinbergAtkinson
Error passed onAll of it (16/16)Three quarters (6/8)
Neighbors46, reaching two pixels right and two rows down
Weights7, 3, 5, 1 out of 161/8 each
Average tonePreservedDrifts toward the extremes
Light and dark areasKeeps sparse dotsClips to clean white or black
Overall lookSmoother, more faithfulCrisper, higher contrast
OriginFloyd and Steinberg's error diffusionBill Atkinson, for the original Macintosh

The important line is the first one. As Wikipedia’s entry on Atkinson dithering puts it, only three quarters of the error is diffused, which weakens results in near-white and near-black areas even though the added contrast can look better. Beyond Loom’s comparison of the two reaches a similar view: Atkinson seems to give richer contrast, at the cost of detail in very light and very dark regions.

What happened on the Spectra 6 palette

Most comparisons use black-and-white images, so we ran both kernels against the six-color palette from our panel definitions. Each row is a flat 120 × 120 patch of one color, dithered with a one-way scan and nearest-color matching in sRGB. We dropped the first 20 rows and columns so edge effects did not count, then averaged the output pixels.

InputFloyd-Steinberg averageAtkinson averageFS: most common pixelAtkinson: most common pixel
Gray 245245, 245, 245255, 255, 25594.0%100%
Gray 230230, 230, 230255, 255, 25584.2%100%
Gray 128128, 128, 128125, 129, 12527.5%28.5%
Gray 2525, 25, 250, 8, 063.5%95.8%
Gray 1010, 10, 100, 0, 085.4%100%
Skin tone (224, 172, 138)224, 172, 138245, 179, 14052.0%55.0%
Orange (240, 130, 30)238, 130, 30238, 131, 048.9%51.5%

How to read it

Limits of this test. These are synthetic flat patches measured in software. Averages are in sRGB values, the palette uses our nominal panel anchors, and no physical panel was photographed. It shows what the math does to tone. It does not show which image a person prefers, and that is a separate question.

Why we still ship Atkinson

Accuracy did not decide our default. Our browser upload pipeline uses plain Atkinson in sRGB with no tone adjustment, because that pipeline came first in a preference ranking we ran across candidate image pipelines, with a mean score of +4.22 against +2.63 for the next best.

Two cautions apply. That ranking compared whole pipelines on our own images and our own preferences, so it is not a clean head-to-head of two kernels and may not match your content. And clipped highlights are a real loss on some images: a pale product on a white background can lose its outline. Preference is a good default for mixed uploads, while a brand team approving a specific image should look at the dithered output itself.

Which dither to use for which content

A product does not need to pick one forever. Our Python optimizer selects by content type. The table below follows its presets, with Floyd-Steinberg added as the other full-error option for photos.

ContentDitherWhy
Text, room signs, status boardsNo ditherKeeps letter edges clean
Line art, logos, flat illustrationAtkinsonClean fills and crisp lines
Bold posters and promo graphicsOrdered (Bayer)Our optimizer's poster preset; a regular pattern and sharp edges
Photos with soft gradientsStucki or Floyd-SteinbergFull-error kernels keep tone
Mixed user uploads, no human tuningAtkinsonRanked first in our preference test across many images

Business signs are often mostly text, which makes the first row the most important. A menu board or meeting-room sign should render its type in solid pigment and reserve dithering for any image on it.

Three practical checks

If you are choosing hardware as well as software, our comparison of Spectra 6 and other color e-paper and our guide to 7.3-inch vs 13.3-inch panels cover the decisions that sit upstream of the dither.

FAQ

What is the difference between Floyd-Steinberg and Atkinson dithering?

Both push each pixel's rounding error onto neighboring pixels. Floyd-Steinberg passes on all of the error to four neighbors, in shares of 7/16, 3/16, 5/16 and 1/16. Atkinson gives 1/8 to each of six neighbors, so only 3/4 of the error is passed on and the rest is dropped. Floyd-Steinberg preserves average tone better; Atkinson produces more contrast and cleaner light and dark areas.

Is Atkinson or Floyd-Steinberg better for e-paper?

It depends on the content. In our flat-color test on the Spectra 6 palette, Floyd-Steinberg reproduced average color almost exactly, while Atkinson turned light grays to pure white and near-blacks to black. That makes Floyd-Steinberg the safer choice for tonal accuracy and Atkinson attractive for illustration, line art and images where clean backgrounds matter.

Why does Atkinson dithering blow out highlights?

Because it discards a quarter of the error at every pixel. In a light gray area the small errors that would eventually add up to a dark dot are partly thrown away, so the dot never appears. In our test, a flat gray of 230 out of 255 came out as 100% white pixels under Atkinson and about 84% white under Floyd-Steinberg.

Which dithering algorithm is best for Spectra 6 photos?

Our optimizer's photo preset uses Stucki, a wider error-diffusion kernel that passes on all of the error. Our general browser upload path uses Atkinson because it ranked first in our preference testing. If skin tones or soft gradients matter, compare a full-error kernel against Atkinson on the real panel.

Should text be dithered on e-paper?

No. Dithering breaks up letter edges. Render text and UI in solid panel colors with no dither, and if a sign mixes text and photos, dither only the photo region.

What is serpentine scanning in dithering?

Processing alternate rows in opposite directions, which reduces the directional patterns a one-way scan can leave. Our Python optimizer scans in a serpentine pattern. Our browser pipeline deliberately does not, because the non-serpentine version is the one that won our preference ranking.