I have long been fascinated by the way e-paper creates images with reflective pigments instead of glowing pixels. Spectra 6 was my first color e-paper display, and I wanted to understand how it made color and whether I could improve the result.
I began with an experimental ICC profile, a color-management description of a fixed rendering-and-display setup, applied before Pimoroni's image conversion. While evaluating the display, I found something unexpected: rearranging the same native colors could change the measured output. That surprise led to controlled pattern tests across pixel arrangement, patterned extent, presentation history and image surroundings. Understanding those measured behaviors is the center of this investigation; prediction-guided rendering is a future direction, with reliable prediction and better image reproduction still open questions.
Most conventional RGB display systems create a requested color by varying the red, green, and blue channel output within each pixel. At the controller interface used in this project, a Spectra 6 pixel is instead assigned one of six native color states: black, white, red, yellow, green, or blue.
To create colors between those native states, a renderer distributes them across neighboring pixels. This is spatial color mixing, often called halftoning or dithering. From a normal viewing distance, the small pattern can blend into an intermediate color. There is no native cyan or magenta state at this controller interface; those and other intermediate colors depend on spatial mixing.
The renderer still has to choose an exact pixel arrangement. That makes the pattern part of the color-generation problem, not merely a decorative detail.
Why arrangement matters: spatial arrangement can change measured color, so useful forecasts may need more than color proportions. We can also compare rendering methods directly on images while prediction remains incomplete. An experimental ICC profile describes a declared rendering/display condition; its usefulness and limits still need measurement.
Two questions guide the rendering work: can we forecast the color a pattern will produce? And can we choose pixels that bring it closer to the desired color?
For the physical display technology, see E Ink's explanation of how electronic ink works, including the four-particle system used by Spectra 6. The pigment system inside the screen and the six color states available to this project's image renderer are different layers.
Enlarged pixel-level teaching diagram. The checkerboard uses two of the panel's six native states to show spatial mixing; measured appearance is evaluated later from spectra and XYZ.
The hardware I used is the Pimoroni Inky Impression 7.3-inch, a color e-paper display for DIY electronics projects that connects to a Raspberry Pi. It combines E Ink Spectra 6 panel technology with Pimoroni's board and Python software and image examples.
Pimoroni's image-conversion software is the starting renderer here. It takes an ordinary RGB image, chooses among the six native states, and distributes those choices across neighboring pixels using dithering. I could therefore begin with a working image display and ask whether measured color information could improve its choices.
Modules like this let people build their own picture frames, signs, and other devices without designing the display electronics from scratch. Other suppliers, including Waveshare and Good Display, also provide color e-paper hardware with example code. Their software and image-preparation routes differ; the Pimoroni route tested here is one specific implementation, not a result for every Spectra 6 product.
My first experiment kept Pimoroni's renderer and added color management before it. Here, the experimental ICC profile described the color response of the rendering-and-display setup under its characterization conditions, so a color-management system could translate image colors for that fixed process. I used it to convert the source sRGB image into new RGB values intended for this display, then passed those values to the same Pimoroni renderer.
Renderer outputs for Vincent van Gogh's Cypresses, from The Metropolitan Museum of Art. The stock and ICC-managed views are the generated six-state pixel maps, displayed here with this project's measured native-state reference colors. The 133-color comparison below supplies the optical measurement of the two routes.
I could also see the difference when displaying the images on the physical panel. That is a visual observation; the numerical comparison below comes from a separate color grid.
Vincent van Gogh, The Bedroom, Art Institute of Chicago. The stock and ICC-managed views are generated six-state pixel maps displayed with the same measured native-state reference colors.
The generated pixel maps show which native-state choices changed. A separate test measured 133 source colors to compare the routes numerically. Lower DeltaE00 means closer to the requested color. Under that campaign's conditions, the dense ICC-managed route had a lower average color error than the stock route.
| Rendering route | Mean error | Largest error |
|---|---|---|
| Stock renderer | 19.9922 | 42.1941 |
| Sparse ICC + renderer | 12.2568 | 26.4508 |
| Dense ICC + renderer | 11.2006 | 23.7069 |
The optical comparison used a separate 133-color grid rather than the paintings shown above. Errors used measured-media-relative D50 Lab against each requested target, averaging repeated errors per source and then the 133 source values equally. The dense route had the lower overall mean while individual colors varied. This development comparison applies to the tested renderer, display and measurement condition.
“Sparse” and “dense” refer to the color grids used to construct the profiles, not the display's pixel density. The dense version used a modeled grid built from measured native states and an optical mixture model. The artwork route used relative-colorimetric conversion before the unchanged palette separation. This early Argyll-based experiment is distinct from the device-neutral ICC profile-writing tool now being developed.
Adding the ICC conversion changed the pixels Pimoroni chose and reduced average error on the measured grid. To improve pixel selection further, I continued measuring the display's color response. Those measurements exposed an unexpected dependency: keeping the color proportions the same did not always keep the measured color the same.
The first clear comparison used blue/green pixels: horizontal rows and vertical columns with the same counts gave different readings. I checked the digital patterns, repeated the measurements, and varied their construction. H/V stripes became a useful diagnostic, not the limit of the investigation. The wider question was which pixel arrangements a renderer could use to produce predictable color.
A pixel payload is the digital data that tells the display which native color state to use at each pixel position. It specifies the pixel arrangement, not the color of the reflected light. The driver packs those state choices into bytes for the display.
A simple area-count description would treat these two patterns alike. These enlarged grids reproduce the repeating row and column rules used in the test. Each full 800×480 pixel map contains 192,000 blue and 192,000 green pixels. Only their spatial organization differs. The measured spectrum and color coordinates below show the optical result.
B = blue; G = green. The letters label native states; the actual payload is encoded as bytes. These grids illustrate the pixel instructions, not predicted appearance or photographs of the panel.
Blue/green prompted the investigation. Later, a one-pixel black/yellow comparison showed a much larger lightness difference. The two curves show how much light each horizontal and vertical pattern reflected at each wavelength. The lower curve shows their wavelength-by-wavelength difference. The accompanying color coordinates and distances describe different parts of the same result.
This interactive plot requires the measurement data. The findings and example values remain available in the text.
One exact one-pixel black/yellow comparison, measured twice in each direction. The interactive atlas shows all four readings.
The native colors and their proportions stayed the same; their positions changed. Some rearrangements gave different readings and others stayed close. Counting colors alone could not describe all the results. To understand why, I needed to separate the pixel instructions from what the display produced and what the meter collected.
An exact pixel map is an instruction to a physical display, not a measurement of the color it will produce. The panel must form the image, and the meter samples the light it returns. A prediction is a forecast of that measured result, not another set of pixel instructions. These distinctions led to spatially informed prediction tests and comparisons of rendering methods.
Grouping pixels reduced some stripe differences, but it did not consistently make arrangements measure alike: a separate four-color comparison gave a larger difference with larger blocks. See that comparison.
The black/yellow example combines two questions: how did the reflected spectrum change, and how did the resulting color change? The blue/green and red/green examples below show why their answers need not track one another. Spectral RMS is useful because it summarizes the average size of a difference across the measured wavelengths. It does not say what either color was, which way the color moved, how far apart the colors are perceptually, or whether a model predicted either one.
RMS means root mean square. At each wavelength, the vertical mean spectrum is subtracted from the horizontal mean spectrum. RMS squares those signed differences so positive and negative regions cannot cancel, averages the squared values, and takes the square root. The result is therefore an unsigned size in percentage points. A value such as 1.58 pp RMS is not automatically large or small; it needs the absolute colors, signed spectral shape, DeltaE00 color distance, and repeated-reading movement beside it.
Each curve is the mean of two readings from one four-reading comparison. The same native plotting style is used later in the interactive atlas, where all four individual spectra can also be inspected.
Higher spectrum difference, smaller color distance
This interactive plot requires the measurement data. The findings and example values remain available in the text.
Lower spectrum difference, much larger color distance
This interactive plot requires the measurement data. The findings and example values remain available in the text.
Where the horizontal and vertical patterns landed in standard color coordinates. These are the values a future color model must predict. CIE colorimetry recommendations define the underlying conventions.
Which way lightness, chromatic components and each wavelength moved. Direction disappears in an unsigned magnitude.
How far apart the two measured colors are under the stated white-reference convention. It is a distance, not automatically a visibility threshold. See Luo, Cui and Rigg (2001) for the development of CIEDE2000.
The unsigned root-mean-square size of the wavelength-by-wavelength difference. Useful, but not the color itself and not a direction.
Why the distinction matters: in the all-pair study, a one-pixel red/green pattern had lower spectral RMS than blue/green (1.58 versus 1.97 percentage points) but a much larger DeltaE00 color distance (15.19 versus 3.31). This does not rank the pairs; it shows that spectral magnitude and perceptual color distance answer different questions. Explore every measured pair and width.
Error diffusion: R. W. Floyd and L. Steinberg (1976), An Adaptive Algorithm for Spatial Greyscale, Proceedings of the Society for Information Display 17(2), 75–77. This is the algorithm's original publication; the particular Pimoroni/Pillow path is a software implementation inspected for this project.
Color difference: M. R. Luo, G. Cui and B. Rigg (2001), The development of the CIE 2000 colour-difference formula: CIEDE2000, Color Research & Application 26(5), 340–350. The paper develops the color-distance formula used here.
Implementation: G. Sharma, W. Wu and E. N. Dalal (2005), The CIEDE2000 Color-Difference Formula: Implementation Notes, Supplementary Test Data, and Mathematical Observations, Color Research & Application 30(1), 21–30. The authors provide test data and discuss numerical pitfalls.
Color coordinates and reference light: CIE 015:2018, Colorimetry, fourth edition, covers standard observers, illuminants, XYZ, color spaces and color differences. The CIE also publishes its D50 spectral data.
These references explain the measurement language. The pattern comparisons, forecast outcomes, and chosen error limits on this page come from this project's experiments.
An ordinary photograph starts as RGB values. Pimoroni's image converter chooses how to represent those colors using the display's six native states, including how to spread them across neighboring pixels. That made image conversion an early suspect.
For the controlled pattern tests, I chose the pixels directly. I specified the native color at every pixel instead of asking Pimoroni to choose a dithered pattern from an ordinary RGB image. I then checked that the pixel buffer produced by the vendor software matched the intended state map and color counts before displaying it. Because the saved buffer matched the pattern I had designed, the unexpected differences were not caused by the image converter choosing different pixel colors or rearranging them.
See the tested pixel arrangements below.
The Pimoroni software and hardware still packed, sent, and refreshed those exact patterns on the display. This check isolated the image-conversion choice, not every part of the system. The source used exact entries from the driver's palette; its set_image conversion still ran, but its output was checked to leave the intended native-state map unchanged. Native-state packing, the low-level driver, the refresh sequence and the panel itself were still involved. Verifying the sent pixels does not establish that every physical pixel reached its intended optical state.
The first meter, an X-Rite i1Pro 3, touches the panel and supplies its own light. Could that contact or illumination contribute to the unexpected readings? To check, I measured selected exact patterns with a Colorimetry Research CR-250, which observes the display from a distance under external illumination.
A second instrument saw selected pattern changes too. For selected patterns, the spectra changed in a similar direction with the non-contact CR-250 and the i1Pro 3. That makes it less likely that the difference came only from the i1Pro's calculation or readout. But the meters used different spots and lighting, and timing and display history could also matter. The comparison does not rule out the i1Pro's light, contact or sampling of the pattern contributing to its reading, and it does not identify the physical cause.
The instruments report different kinds of measurements, so I compared whether the spectra changed in the same direction, not whether their numerical values matched.
CR-250 blue/green readings → i1Pro blue/green readings →
To follow changes over time and after another display update, I recorded 108 CR-250 observations alongside i1Pro readings. The CR-250 viewed a different exposed spot on the same display presentation. Some sequences put an i1Pro reading between two CR-250 readings without refreshing the display.
CR-250 hold/redraw sequence → · i1Pro hold/redraw sequence →
The i1Pro 3 used here is the standard model. X-Rite's specification sheet lists its nominal 4.5 mm measurement aperture. At about 0.20 mm per display pixel, that diameter spans roughly 22.5 pixels. The meter combines light from many pixels. Moving it slightly or shifting the pattern can change which pixels contribute most to the reading.
Simulations that treated the spot as a circle and counted the pixels inside it equally did not explain the observations by themselves. They also did not measure how strongly different parts of the real spot contribute to the meter's signal.
Already measured: CR-250 → i1Pro 3 → CR-250 sequences on an unchanged display presentation. The i1Pro was already touching the panel, and the CR-250 observed another exposed spot.
Still an open comparison: repeat that sequence alongside a matched waiting period with no i1Pro reading. That would help distinguish change associated with using the i1Pro from change that also happens while waiting. The completed interleaving does not provide that matched control. Placing or moving the contact instrument would be a separate test; none of these comparisons alone identifies a single physical mechanism.
The CR-250's radiance and the i1Pro 3's reflectance factor are different quantities. Cross-route comparisons here concern spectral direction, not pooled numerical magnitudes.
I automated the display updates and spectral readings. Each four-reading sequence measured both patterns twice. For selected comparisons, I also swapped which pattern came first and last. This let me see both whether the difference returned and whether either pattern changed between its two readings.
Each pattern is read twice. Putting one pattern first and last, then swapping their positions, helps reveal changes associated with reading order or time. It does not make changes over time or effects of earlier images disappear.
A and B label the two patterns being compared. In this stripe example, A is horizontal and B is vertical. In other comparisons both patterns can be horizontal; the letters still describe reading order, not the pattern's orientation.
Each saved spectrum is tied to the exact source image, packed native buffer, instrument settings, presentation time, and code version. SHA-256 file fingerprints let me verify that repeated constructions really use identical pixels. Raw spectral files and the complete XYZ/Lab results remain available alongside the summaries.
The stripe-width study alone collected 300 optical readings of 90 exact native pixel patterns; later experiments expanded into multicolor layouts, presentation history and image context.
The H/V discovery led to tests of checker layouts, multicolor organization, ordered versus error-diffused white/blue/green pixels, starting position (phase), patterned extent, presentation history and image surroundings. Regrouping three colors changed the measured result while every row stayed horizontal and the proportions stayed fixed. Some rearrangements produced large color differences; others produced small ones. The color pair and the layout both mattered.
The visual map covers 11 investigated situations. H/V stripes are one family, with all 15 native-color pairs available inside it. The results show what changed when pixels were rearranged, held on screen, or placed within a picture.
The same amounts of native color sometimes measured very differently after rearrangement; other comparisons stayed close. These tests show which was which and whether the contrast appeared again.
If only color counts mattered
Patterns with the same amounts of each native color would measure the same regardless of where those pixels were placed or what had been displayed before. That simple starting idea assumes each native color reflects consistently and the meter combines light from the pixels in the same way in every arrangement.
The previews show the actual digital pixel maps, not photographs or measured-color swatches. Fine patterns are enlarged from exact 48 × 32 pixel crops; select one to inspect its full 800 × 480 frame. Inlay and image-context previews show the full frame so their surroundings remain visible. ΔE00 compares the measured colors; RMS compares their spectra. Y is the measured light-level coordinate within XYZ, not perceived brightness. How these measurements describe color.
18.24 ΔE00
0.35 ΔE00
The partner mattered: yellow-containing pairs were prominent, but not uniformly affected. Horizontal was not always lighter.
On repeating: Blue/green’s color gap returned at 3.21–3.39 ΔE00 across four four-reading sequences with the H/V order reversed. A small black/white difference changed spectral direction between sessions. Even when the gap repeats, the two patterns can both have shifted in color.
All 15 comparisons ↓ Large black/yellow → Small white/yellow difference → Blue/green repeat sequence → Black/white context → Width / reverse order →
This complete orientation screen covers every pair of the display's six native colors. Each row compares one-pixel horizontal and vertical stripes with equal color counts. H and V each summarize two i1Pro readings in the stated order. Select a pair to open its four spectra and exact measurement records.
The pair changed the size of the H/V difference. Black/yellow and red/green had much larger color separations than white/yellow and white/blue. Yellow-containing pairs were prominent, but the partner controlled the result, and horizontal was not always lighter. Pairs that stayed small under this test are not universally unaffected: a different construction or measurement condition is a separate question.
Yellow paired with black, blue, red or green gave the four largest spectral RMS differences in this one-pixel set. Color-distance rankings were different: red/green was second by DeltaE00 but sixth by RMS. These describe the commanded color combinations, not a finding that yellow pigment caused the differences. The table shows both rankings, with 1 meaning the largest value in that column.
Swipe or scroll to see all columns →
| Native pair · digital stripes | H/V measurement sequence | Measured color previews | H–V color distance ΔE00 | Spectral RMS pp | Lightness H−V ΔL* | Same-pattern movement pp RMS |
|---|---|---|---|---|---|---|
| Interactive pair details require the measurement data. | ||||||
The stripe miniatures illustrate digital pixel instructions, not measured appearance. The separate H/V swatches are approximate sRGB previews derived from measured XYZ. Positive ΔL* means H measured lighter; negative means V did. Same-pattern movement compares the two H readings and the two V readings separately, not an uncertainty interval or pass/fail boundary.
A selected one-pixel comparison gave 3.31 ΔE00. Two-pixel comparisons gave 0.15 ΔE00 and 0.95 ΔE00 in opposite orders.
On repeating: The gap was smaller with wider stripes here, but its size changed when the reading order was reversed. These results do not show that wider stripes will make horizontal and vertical patterns match reliably.
H/V separations: 0.045 ΔE00, 0.090 ΔE00, and 0.008 ΔE00.
On repeating: Each pattern was measured twice. Rows and columns measured close to each other at all three yellow percentages. That does not yet tell us how to predict the color of any white/yellow mixture, or whether it will return to the same color later.
2.32 ΔE00 between mean colors. The checker moved 3.82 ΔE00 between its two readings.
On repeating: The checker changed more between its two readings than the average checker differed from the average stripe. That makes it hard to attribute their difference to layout alone.
0.13 ΔE00; 0.047 pp RMS.
On repeating: For both layouts, the spectrum changed more between repeated readings than it differed between their averages. The small average gap alone does not show that the two layouts will consistently produce the same color.
Grouping: 15.39 ΔE00; 12.083 pp RMS. A separate test of row order, with the lengths of same-color bands matched: 0.39 ΔE00.
On repeating: The exact cyclic/grouped contrast returned at 15.57 ΔE00 in reverse order; the separate row-order contrast also recurred. Grouping also changed the lengths of same-color bands and where colors met, so the test does not tell us which of those changes mattered.
0.14–0.20 ΔE00 across four i1Pro comparisons.
On repeating: Some readings changed as much or more when the same pattern was measured again than when its layout changed. The small gaps therefore do not establish that the two layouts always produce the same color.
0.38 ΔE00; 0.242 pp RMS in the selected order.
On repeating: Reversing the reading order produced a difference in the same direction. This tests one starting-position shift, not every possible alignment.
Complete digital frames above, reduced by area averaging to show the changing surround; enlarged stripe crops below. Neither is a measured-color preview.
H/V differences inside the small squares: 0.10–0.29 ΔE00; across the full screen: 3.13–3.61 ΔE00, across both orders.
On repeating: All three sizes were measured in both reading orders. Enlarging the colored area also shrank its white surround and moved its edges. We cannot tell from this comparison which change caused the larger gap.
Would the same pixels measure the same color inside a picture and against a white background? The outlined 64 × 64 pixels below are identical in both images; only their surroundings change.
These are the exact digital pixel maps sent to the display, not photographs or simulations of the measured colors. The outline marks the pixels kept unchanged, not the meter's measurement area.
The average measured colors differed by 3.23 ΔE00.
On repeating: We measured the image twice and the white-background version twice. Comparing each image reading with each white-background reading gives four pairings, with a largest difference of 3.33 ΔE00. Those pairings reuse the same four readings; they are not four separate experiments.
Why the cause is still open: The meter collects light from an area, not a single pixel. We measured the patterns, but did not map how much each part of the screen contributed to its signal. This comparison therefore cannot tell us whether the difference came from the surrounding image changing the display's response, the way the meter collected light, or changes during the measurement sequence.
The measured Y value fell after another redraw but rose slightly while held, after each of four preceding images.
On repeating: That direction recurred across the four preceding images, but each hold and redraw episode was observed once and started at a different level. This does not establish a general drift rule or isolate a cause.
The identical pale-red pattern did not return to one measured color.
On repeating: The pattern was measured twice at each early, middle and late return. This sequence does not isolate elapsed time from the intervening display history.
The pale-red pixel buffer was identical at every return. The timeline shows those six readings.
A separate test changed the preceding image; it is not part of this pale-red return sequence. Inspect the preceding-image comparison →
The labels summarize these measurements; they do not say whether a difference would be visible to a viewer. A small difference in one test is not a promise that the patterns will always match.
A recurring difference does not mean that either pattern returns to the same measured color. Some differences repeated, while other readings changed with time or surroundings. These findings motivate testing whether spatial arrangement, surrounding pixels and display history improve forecasts. A future renderer would use that knowledge to reproduce requested colors better. Three open questions guide that next direction.
Wider stripes usually reduced the spectral contrast in this set. For 14 of 15 pairs, the arithmetic mean of the two two-pixel H/V comparisons was below the one-pixel RMS; the two three-pixel comparisons gave the same count. Phase, frame boundaries, and the pattern inside the meter's spot also changed, so width alone is not a general fix.
The color table above compares one-pixel patterns. This plot compares the size of the spectral H/V difference across all three stripe widths and both reading orders. It shows where contrasts became smaller or differed between orders; it does not measure perceptual color distance or prediction accuracy. Equal spacing on the logarithmic axis represents equal ratios, keeping small and large contrasts readable together.
Keyboard: use arrow keys to inspect comparisons; Home and End jump to the first and last. Tab leaves the plot.
This interactive plot requires the measurement data. The findings and example values remain available in the text.
Interactive history details require the measurement data.
Interactive history details require the measurement data.
A displayed image could remain fairly steady while held, yet return to a different color after another update or setup. The direction depended on the pattern. That made presentation history part of the prediction question, rather than a single correction to apply to every color.
Tests with different preceding images and an extra white update did not reveal a consistent clearing procedure. One proposed reset sequence, blue then white before returning to the same white/blue/green pattern, changed the reading in the opposite direction to the intended correction. Those results narrowed which remedies were worth pursuing; they did not rule out every history effect.
The experiments compared changes within held episodes with changes after another refresh, and also tested returns after different preceding images. The extra-white test did not consistently improve the result, and the white control itself changed on return. A curve fitted to the recorded time trend can describe those readings, but it is not a prediction made before measuring them. Normal display use is still one update followed by a held image.
These comparisons extend the same investigation beyond horizontal versus vertical. They keep the patterns horizontal and ask how the measured color changes when the row sequence, starting position, and run grouping change. Together, they show that orientation was only one way exact organization could change the measurement.
This comparison kept both patterns horizontal and used the same 96,000 black, 192,000 yellow, and 96,000 blue pixels. One four-row period alternated yellow–blue–yellow–black; the other placed the two yellow rows together as yellow–yellow–blue–black. The display received two exact pixel arrangements, not two names for one image.
The cyclic pattern measured lighter and more yellow in both presentation orders. Its signed Lab change was ΔL* +15.39 / +15.51 and Δb* +11.04 / +12.04.
DeltaE00 was 15.390 in the first order and 15.573 in the reversed order. This is an unsigned separation under nominal perfect-diffuser D50 convention. It describes color separation, not a visibility or significance verdict.
All 36 measured bands pointed in the same signed direction in both orders. Spectral RMS was 12.083 and 12.200 percentage points.
The two signed spectral shapes had cosine 0.999900; their vector discrepancy was 0.207 percentage points RMS. A cosine close to 1 means the spectral changes point in nearly the same direction, even if their sizes differ. Vector discrepancy is the remaining difference between those two change spectra. Movement between repeat readings of the same pattern remains context, not an uncertainty limit.
Six four-reading comparisons separate three questions about organization. Every pattern kept the same black/yellow/blue counts and the same horizontal axis. A run is a band of consecutive rows with the same color. Boundary density describes how often neighboring rows change color. Each row remains its own descriptive result because it represents a different exact comparison; they are not averaged into one rule.
Swipe or scroll to see all columns →
| Arrangement question | Order | Color distance DeltaE00 | Spectrum difference pp RMS | Same-pattern movement first / second, pp RMS |
|---|---|---|---|---|
| Move the starting row of the cyclic sequence | ABBA | 0.379 | 0.2425 | 0.1717 / 0.1051 |
| Move the starting row of the cyclic sequence | BAAB | 0.293 | 0.2110 | 0.0705 / 0.1298 |
| Change cyclic rows into two-row runsRun length and boundary density change together | ABBA | 15.626 | 12.2002 | 0.2227 / 0.0946 |
| Change cyclic rows into two-row runsRun length and boundary density change together | BAAB | 15.596 | 12.1601 | 0.0851 / 0.0247 |
| Compare two arrangements with the same simple summariesSame color counts, same number of boundaries, same counts of each run length, and same counts of which color directly followed which | ABBA | 0.385 | 0.3672 | 0.0992 / 0.0530 |
| Compare two arrangements with the same simple summariesSame color counts, same number of boundaries, same counts of each run length, and same counts of which color directly followed which | BAAB | 0.412 | 0.3722 | 0.0281 / 0.0657 |
In this exact family, the 0.2425 / 0.2110 pp starting-phase differences were smaller than the large arrangement and run-grouping contrasts, but each was above both duplicate movements in its order (0.1717 / 0.1051 pp and 0.0705 / 0.1298 pp). Their signed directions also recurred across orders (cosine 0.9896). This does not predict what other starting positions would do or identify the physical cause.
The large 12.2002 / 12.1601 pp difference recurred with nearly identical signed spectral direction (cosine 0.999986). Run length and boundary density changed together, so neither is isolated as the cause.
The smaller 0.3672 / 0.3722 pp difference also recurred in direction (cosine 0.995871). Matching those summaries did not make the exact arrangements measure identically.
The comparison set repeats the cyclic-versus-two-row comparison, then compares two-row and four-row runs in both presentation orders. The exact black/yellow/blue counts and horizontal axis stayed fixed, while run length and boundary density still changed together.
Swipe or scroll to see all columns →
| Arrangement question | Order | Color distance DeltaE00 | Spectrum difference pp RMS | Same-pattern movement first / second, pp RMS |
|---|---|---|---|---|
| Compare cyclic rows with two-row runsRepeated alongside the longer-run comparisons in the same measurement sequence | ABBA | 15.588 | 12.1619 | 0.0827 / 0.0969 |
| Compare two-row runs with four-row runsRun length and boundary density change together | ABBA | 2.257 | 0.8491 | 0.0528 / 0.0555 |
| Compare two-row runs with four-row runsRun length and boundary density change together | BAAB | 2.290 | 0.8591 | 0.0464 / 0.0390 |
The 12.1619 pp cyclic-versus-two-row comparison pointed in almost the same signed spectral direction as the corresponding ABBA and BAAB rows (cosines 0.999992 and 0.999997).
The two-row-to-four-row comparisons were 0.8491 and 0.8591 pp with cosine 0.99982394, magnitude ratio 1.01183170, and signed-vector discrepancy 0.01891485 pp RMS.
The two-row-to-four-row difference was about 7% of the cyclic-to-two-row difference's RMS and its spectral shape differed (cosine about 0.44). Run length and boundary density changed together; these observations do not define a proportional scaling law.
A separate comparison used equal amounts of white, black, green and blue. It compared a regularly repeating arrangement with an irregular, computer-generated arrangement at two block sizes. With 2 × 2 same-color blocks, their measured colors were close; with 4 × 4 blocks, the difference was larger in both reading orders.
| Same-color block | ABBA | BAAB |
|---|---|---|
| 2 × 2 pixels | 0.226 | 0.123 |
| 4 × 4 pixels | 2.241 | 2.172 |
Enlarging the blocks also changed local color proportions and edge positions within the nominal measurement area; the meter’s exact spatial weighting was not established. This result argues against assuming that larger blocks always reduce arrangement-dependent differences, but it does not isolate block size as the cause.
Prediction-guided rendering is work in progress, not a completed solution. The measured behavior raises three open questions: which patterns remain predictable, whether that knowledge transfers into images, and whether it helps choose pixels that reproduce requested colors better.
Forecast a construction's output, including when the same pixels return.
Test the color of a region among the textures, edges and surroundings of a picture.
Choose native pixels that improve color while preserving image structure.
Different models can predict different colors for the same unchanged payload. Making a forecast does not change the pixels. The requested image color is a separate goal: we check both how well we predicted the output and how closely that output matches the goal.
In questions 1 and 2, prediction error means the DeltaE00 distance between a saved forecast and a fresh measured color. Those two tests chose a limit of 1 before measurement, using nominal-D50 Lab. That is an engineering criterion for those tests, not a universal visibility threshold. Question 3 also checks distance from the requested color, labeled separately.
A working hypothesis about possible model inputs, not a validated recipe. Which arrangement and condition terms matter remains open. The empirical forecasts discussed here predict XYZ directly.
The direct model averages measured XYZ from nearby inputs. The ratio model averages how those measurements differ proportionally from a native-color estimate, then applies that ratio to the new input's estimate. The two approaches were compared with forecasts saved before the new readings. Neither changes the displayed pixels by itself.
The practical goal is to use what we learn about the display to choose native pixels that bring measured colors closer to the image's requested colors, while preserving tone, detail, texture and edges.
In an earlier two-color experiment, I used previous measurements to adjust the renderer's input for a warm tone and a neutral gray. This was a fixed, case-specific adjustment, not a general predictor choosing pixels.
| Test color | Baseline rendering | Adjusted rendering |
|---|---|---|
| Warm tone | 6.24–6.65 | 2.86–2.98Closer on every reading |
| Neutral gray | 5.51–5.87 | 6.17–6.51Farther on every reading |
The adjustment did not improve both colors. Its saved output forecasts also missed the chosen limit on all eight readings. Every adjusted result remained above the test's original-color error limit of 1. The warm improvement was useful, but it did not establish a dependable correction for other colors.
Both versions were compared with the same original desired color for each case. These patterned test patches did not test whole-image quality or establish the closest colors the display could reproduce. Inspect the readings and forecasts.
A separate construction in that experiment had an accurate output forecast but still missed the desired color. Its forecast was saved before four fresh measurements; the first reading is shown here.
All four fresh readings told the same bounded story: forecast error was 0.38–0.47 DeltaE00, while requested-color error was 5.81–5.96.
| Reading | Measured D50 XYZ | Forecast error | Requested-color error |
|---|---|---|---|
| 1 | 14.29114 · 12.57847 · 5.426062 | 0.4089 | 5.9641 |
| 2 | 14.24502 · 12.54802 · 5.422607 | 0.3914 | 5.9523 |
| 3 | 14.39455 · 12.68825 · 5.477240 | 0.4711 | 5.8099 |
| 4 | 14.28953 · 12.60059 · 5.436232 | 0.3792 | 5.8571 |
For prediction-guided rendering, forecast accuracy and requested-color accuracy answer different questions. Dependable prediction-guided reproduction across new images has not yet been demonstrated.
What comes next? Test whether a selected correction helps on fresh colors and images, with its forecasts fixed in advance. Evaluate requested-color error alongside tone, detail, texture and edges.
What began as an experimental profile evaluation became a test of how exact pixels produce color. Repeated controlled comparisons reproduced differences, and fixed-axis three-color patterns showed that orientation was only one part of the problem. The current aim is better color reproduction, through parallel work on prediction and direct image comparisons. We can test which constructions are predictable, whether forecasts transfer into images, and which rendering methods produce more useful pictures. Image comparisons do not have to wait for prediction to succeed.
The first experimental ICC profile characterized an earlier rendering-and-display setup. A later experimental profile would describe the color response of a declared candidate process being evaluated. It would not execute the renderer or automatically receive missing spatial, context or display-history inputs. The ICC profile specifications define that interchange format.
The investigation is still in progress. These next tests are planned and depend on access to the display and instruments. They ask three different questions:
Stripes have been measured for all 15 pairs, but checkerboards only for selected pairs. Use those results to identify the missing comparisons, then measure fresh stripe controls alongside the new checker readings so the layouts can be compared under the same conditions.
Test a model that adds information about surrounding pixels to its estimate of the measured region. Its method and forecasts are fixed before the new readings, so the test can show whether that extra information helps on fresh image outputs.
Compare fixed rendering methods on known images. Measure local color error and examine tone, detail, texture and edges. This is a separate test from forecast accuracy: a predictable output can still be the wrong color, and a lower color error can still come with lost detail.
Later comparisons could vary checker size and starting position (phase), rectangular and diagonal layouts, arrangements that change both row and column sequences, color proportions, regular patterns versus error diffusion, and the size of a patterned region within its background.
Three- and four-color tests would also change which color occupies each part of an uneven pattern: a narrow band and a wider band give a color different neighbors. The yellow/blue/yellow/black one-pixel pattern was already measured as both rows and columns. The longer two-row and four-row groups shown here were horizontal; testing their matched column versions remains a more specific question.
An offline experimental foundation uses measured native-color references and diffuses XYZ error while choosing pixels. Models for composite outputs and their usefulness in images remain under investigation; this implementation is not a measured general-image improvement.
| Decision | Pimoroni baseline used here | Experimental direction |
|---|---|---|
| Interpret image color | RGB values, with source-profile conversion upstream. | Explicit source-profile conversion or declared sRGB, then D50 color coordinates. |
| Choose native states | Nominal RGB references and nearest-RGB selection. | Measured references and colorimetric coordinates; validate composite-output models before relying on them. |
| Place pixels | Pillow Floyd–Steinberg error diffusion. | The implemented baseline diffuses XYZ error. Any model-guided replacement must also preserve image detail. |
This display has been slower and harder to characterize than I expected. Each pattern needs an update, time to settle and repeated measurements. That challenge has made the work a rewarding way to connect color science, software and experimental design: understand the output well enough to make better choices when rendering an image.
For a closer technical look, the atlas keeps the individual spectra, color coordinates, forecasts and experiment history. Browse stripe width, proportions, arrangement, surroundings, instruments or prediction; this depth is optional, not a prerequisite for following the three questions.
Open the interactive measurement atlas →The original 75 stripe comparisons remain available by color pair, width and reading order. The topic selector opens the broader experiments and later prediction tests.