Ember: Nightshift / Redshift safe color palettes

Terminal, chart, heatmap, and UI palettes that remain usable before and after aggressive warm-screen filtering.

Ember was created to solve the problem of traditional color palettes having colors that appear indistinguishable once you turn on redshift. With Ember, the primary goal is distinctive colors under various color temperature filters. Only once that's satisfied do we further tweak the redshift-degenerate color channels to improve visual separation during daytime (aka - with no redshift applied).

Ember comes with 4 color palettes:

• 3400K Dark
• 3400K Light
• 2000K Dark
• 1200K Dark

The 3400K palettes are good for general use, and approximately optimized for a Macbook's maximum Nightshift effect.

The 2000K and 1200K palettes are more specialized for true "deep redshift" fans (and astronomers).

One interesting consequence is that the 2000K and 1200K extreme redshift palettes have fewer distinct colors than the 3400K palettes. This is because such extreme redshifts dramatically reduce the size of the perceptual color space available to choose colors from! More on this below...

Each palette is authored in commanded sRGB, verified under its corresponding color temperature's modeled per-channel RGB gains, and exported as terminal themes (Alacritty, iTerm2, Windows Terminal), UI surface roles, categorical chart colors, a 256-sample sequential map, and Matplotlib/CSS/JSON/Python artifacts. Ember does not apply the filter; keep using the one you already have.

The four palettes

3400K Dark

3400K Light

2000K Dark

1200K Dark

These are the exact commanded sRGB values shipped in every export. The deeper the target filter, the fewer color identities a family authors; aliased terminal slots are labeled as aliases instead of posing as additional colors.

With and without redshift

In a terminal

In charts and heatmaps

Why identities disappear

A warm filter is not a tinted overlay: it multiplies red, green, and blue by different gains. At Ember's 1200 K model the blue gain is zero, so colors that differ only in blue produce identical output. The deeper palettes therefore author fewer identities instead of pretending aliased colors remain distinct.

Filtered rows are deterministic signal simulations — commanded sRGB multiplied by each profile's published gains. They are not photographs, calibrated physical color temperatures, or predictions of every display pipeline. Turn off any active warm filter before judging them, or your screen applies the transform a second time.

Choose a profile

Palette	Use it when	Distinct categories
3400k-dark	you want a near-black general-purpose warm theme	6
3400k-light	you use a moderate warm shift on a light surface	6
2000k-dark	you run Redshift near 2000 K	4
1200k-dark	you want an extreme 1200 K stress profile	3

Start with 3400k-dark unless you deliberately run a deeper filter. The 2000 K and 1200 K profiles are dark-only because a filtered light canvas becomes a large orange-red field.

Make Ember work

1. Get the files

For terminal themes, CSS, and the JSON manifest:

git clone https://github.com/carpdiem/ember.git
cd ember

Python users can skip the clone and install directly from GitHub:

python -m pip install "ember-palettes @ git+https://github.com/carpdiem/ember.git"

2. Import a terminal theme

• Alacritty: copy one file from themes/terminal/alacritty/ into your config directory, then import it from alacritty.toml:

  mkdir -p ~/.config/alacritty/themes
  cp themes/terminal/alacritty/2000k-dark.toml ~/.config/alacritty/themes/

  [general]
  import = ["~/.config/alacritty/themes/2000k-dark.toml"]

• iTerm2: open Settings → Profiles → Colors → Color Presets… → Import… and choose a file from themes/terminal/iterm2/.

• Windows Terminal: open Settings → Open JSON file, copy one object from themes/terminal/windows-terminal/ into the root schemes array, then set your profile's colorScheme to its exact name.

The terminal guide has the complete import steps and explains how reduced ANSI banks behave under the deep profiles.

3. Use the UI surface roles

Every family exposes the same ordered roles in JSON, CSS, and the Python surfaces() API:

Role	Intended use
bg_0	base application canvas
bg_1	low-emphasis adjacent region or sidebar
bg_2	ordinary panel or card
bg_3	nested panel, active control, or hover state
bg_4	floating panel, menu, or popover
bg_5	selected row, range, or focused region
fg_0	primary text and essential labels
fg_1	larger supporting text or graphics; not normal-size body text
fg_2	muted, nonessential metadata or decoration

The six backgrounds form a monotonic ladder. bg_0 is always the canvas and bg_5 is the strongest background state: dark families become lighter toward bg_5, while the light family becomes darker toward bg_5.

Matplotlib

import matplotlib.pyplot as plt

from ember import categorical, categorical_norm, encode_categories, sequential, surfaces

palette = "2000k-dark"
ui = surfaces(palette)
labels = ["control", "alpha", "beta", "gamma"]
order = ["control", "alpha", "beta", "gamma"]
category_ids = encode_categories(labels, order, slug=palette)

fig, (points, image) = plt.subplots(1, 2)
fig.patch.set_facecolor(ui["bg_0"])
points.set_facecolor(ui["bg_2"])
image.set_facecolor(ui["bg_3"])
points.scatter(
    [1, 2, 3, 4],
    [1.2, 2.4, 1.8, 3.1],
    c=category_ids,
    cmap=categorical(palette),
    norm=categorical_norm(palette),
)
image.imshow([[0.0, 0.4], [0.7, 1.0]], cmap=sequential(palette))
plt.show()

Pass the palette slug to categorical_norm() and encode_categories() so their capacity checks match the selected family. Sequential maps always expose 256 canonical float samples, independent of the number of categorical colors.

CSS

Load ember.css, then select a family:

<link rel="stylesheet" href="/path/to/ember.css">

<section data-ember-palette="3400k-dark">
  <div class="panel">...</div>
</section>

[data-ember-palette] {
  color: var(--ember-fg-0);
  background: var(--ember-bg-0);
}

.panel {
  background: var(--ember-bg-2);
}

.panel:hover {
  background: var(--ember-bg-3);
}

.popover {
  background: var(--ember-bg-4);
}

.selected {
  background: var(--ember-bg-5);
}

.series-a {
  color: var(--ember-category-one);
}

.heatmap-key {
  background: var(--ember-sequential);
}

CSS exposes eleven representative 8-bit gradient stops. The JSON manifest and Python package preserve all 256 canonical float samples, along with surfaces, categorical colors, ANSI slots, gain profiles, and measured results.

What is the science behind it?

1. Model the signal that reaches the display

Ember approximates a warm display transform as an independent gain on each sRGB channel:

display RGB ≈ commanded RGB × [red gain, green gain, blue gain]

Profile	RGB gains	Basis
3400k	[1.00, 0.74, 0.53]	warm-white engineering surrogate
2000k	[1.0000, 0.5436, 0.0868]	pinned Redshift 2000 K signal LUT
1200k	[1.0000, 0.3094, 0.0000]	pinned Redshift 1200 K signal LUT

At 1200 K, blue contributes nothing to the modeled output. At 2000 K, only 9% survives. Ordinary sRGB distance is therefore a bad proxy for nighttime distinction: two colors can be far apart by day and converge after filtering.

These are software signal models, not calibrated physical color temperatures. A real result also depends on the display, operating system, calibration, brightness, and ambient light.

The JSON manifest also reports sensitivity diagnostics at four ±5% green/blue gain corners for categorical colors, terminal groups, foreground/surface contrast, and sequential spacing. These sampled corners expose nearby model sensitivity; they are not extrema over every point inside a continuous gain box and are not display calibration measurements.

2. Solve the constrained state first

Ember treats day and night as two views of the same commanded color. It does not average their quality into one score, because excellent daytime spacing cannot compensate for a nighttime collision.

1. Set hard transformed targets for contrast, lightness/chroma geometry, and minimum perceptual separation in Oklab.
2. Among the commanded colors that reproduce those targets, choose a moderate-chroma daytime set with strong unfiltered separation.

This reverses the usual workflow. At 1200 K, changing only blue cannot disturb the transformed color, so Ember can use that otherwise lost channel to improve daytime identity. At 2000 K, the same freedom is smaller because a weak blue residual remains. Generated release checks keep every serialized accent within 0.15 ΔEOK of its authored transformed target.

Categorical colors must also clear the complete fg_0 / fg_1 / fg_2 ladder in both states, not merely remain distinct from one another. The 2000 K and 1200 K banks repeat their category-spacing, foreground-clearance, and background-contrast checks at all four sampled gain corners.

3. Keep frequent pixels neutral and reserve color for meaning

Human vision carries fine spatial detail more strongly through luminance than chromatic channels. Dense saturated glyphs and opposing hues are therefore poor places to spend a limited nighttime color gamut. The comparison below shows the practical consequence: pure white becomes a brighter transformed orange than Ember's cream body text, while a daytime dark gray becomes a much larger rust-colored signal than Ember's near-black canvas.

Ember puts most pixels in warm-neutral surfaces, uses cream rather than pure white for body text, and reserves higher chroma for semantic accents. Every foreground-capable terminal accent still clears 4.5:1 contrast against the terminal base background (bg_0) after its target transform. fg_1 and fg_2 remain available for larger supporting text and nonessential metadata.

4. Protect identity with both color and structure

Under the current release gates, Ember supports six categorical identities at 3400 K, four at 2000 K, and three at 1200 K. Deep terminal themes repeat those supported capacities across the sixteen ANSI slots; unsupported names alias deliberately instead of pretending to add another color identity.

An accent may change apparent hue between states; it must remain distinguishable in both. Every terminal bank is evaluated with all three foreground roles so an accent cannot pass by colliding with ordinary, supporting, or muted text. Every foreground trio must remain one ordered warm-neutral ladder rather than three unrelated colors.

Color is still not enough for critical identity. Charts should combine it with direct labels, position, dash pattern, marker shape, or texture:

5. Space continuous maps in the transformed view

Each sequential map begins with a human-chosen earth-tone path. The generator smooths that path in Oklab, measures cumulative distance after the target transform, and resamples it at equal transformed-distance intervals. The 2000 K and 1200 K maps use restrained interior blue-channel adjustments to improve commanded spacing without giving up transformed equidistance, endpoints, or monotonic lightness.

The result is a 256-sample map with strictly monotonic transformed lightness and nearly equal modeled transformed Oklab steps. Release checks also require monotonic daytime lightness and bounded daytime step variation. CSS exposes eleven convenient 8-bit preview stops; JSON and Python carry the complete float samples.

Measured properties and exact release gates

Measured properties

ΔEOK below is Euclidean Oklab distance multiplied by 100. It is an engineering measure used consistently by the generator and tests, not a standardized CIE ΔE formula.

Family	Categories	Day min ΔEOK	Transformed min ΔEOK	Mean / max raw chroma	Transformed L range	Min ANSI contrast
3400K Dark	6	15.00	11.45	0.0971 / 0.1045	0.2206	5.29:1
3400K Light	6	16.73	12.41	0.0988 / 0.1037	0.3068	4.65:1
2000K Dark	4	17.00	12.91	0.0956 / 0.1082	0.1538	4.52:1
1200K Dark	3	20.72	10.25	0.1047 / 0.1107	0.1268	4.55:1

Daytime hue breadth and transformed category/background contrast are separate release gates. Contrast here is for graphical category marks, not small text.

Family	Day minimum hue gap	Target	Transformed category / bg_0	Target
3400K Dark	20.23°	≥ 20°	3.01:1	≥ 3:1
3400K Light	31.75°	≥ 30°	3.03:1	≥ 3:1
2000K Dark	27.76°	≥ 20°	3.05:1	≥ 3:1
1200K Dark	65.03°	≥ 45°	3.12:1	≥ 3:1

Terminal-bank separation includes the complete foreground ladder as well as accent-to-accent comparisons:

Family	Day accent min	Day → fg_0	Day → fg_1	Day → fg_2	Transformed accent min	Transformed → fg_0	Transformed → fg_1	Transformed → fg_2
3400K Dark	10.76	8.67	8.13	11.72	7.31	6.58	5.16	10.18
3400K Light	15.64	9.25	9.47	9.22	11.07	6.47	6.70	7.13
2000K Dark	12.62	13.73	10.86	8.25	7.75	7.64	5.03	4.75
1200K Dark	12.35	9.69	8.96	14.68	4.13	4.49	4.13	11.29

Foreground coherence is independently gated rather than assumed:

Family	fg_0 / fg_1 / fg_2	Day adjacent steps	Transformed adjacent steps	Day / transformed gap ratio	Day / transformed hue span	Max day chroma
3400K Dark	#DDD0B2 / #BDAE93 / #908472	10.41 / 13.79	8.70 / 11.88	0.7568 / 0.7372	9.32° / 2.63°	0.0426
3400K Light	#342F2C / #4D4540 / #665C54	8.75 / 8.52	7.55 / 7.32	0.9734 / 0.9707	0.00° / 1.12°	0.0181
2000K Dark	#EED5AE / #D3BB99 / #AA9D8B	8.04 / 10.44	6.20 / 9.04	0.7886 / 0.7008	2.74° / 2.51°	0.0584
1200K Dark	#FFE5BD / #CBAF89 / #A18C73	16.43 / 11.79	11.07 / 9.16	0.7099 / 0.8177	6.73° / 0.71°	0.0607

Dark-surface measurements use WCAG's sRGB relative-luminance calculation on the exact serialized Hex values. The contrast range covers transformed fg_0 on all six background roles.

Dark family	bg_0	Commanded luminance, bg_0 → bg_5	Transformed fg_0 contrast range
3400K Dark	#090807	0.00247 → 0.02019	6.83–8.52:1
2000K Dark	#070504	0.00162 → 0.01852	5.86–6.98:1
1200K Dark	#060302	0.00108 → 0.01571	5.32–6.08:1

These are digital signal measurements, not physical display luminance. Actual black level still depends on panel technology, brightness, calibration, ambient light, and the display's behavior near black.

The build also checks fg_0 against every declared background, verifies endpoint visibility, parses every terminal format, and reproduces all generated artifacts from source.

Reproduce the build

uv sync --extra dev
uv run python tools/build_all.py --check
uv run pytest -q
uv run ruff check src tests tools examples
uv build

The release gates enforce:

• exactly four palette families with categorical capacities 6, 6, 4, 3;
• categorical commanded mean Oklab chroma between 0.09 and 0.105, with no color above 0.111;
• categorical minimum-distance floors in both unshifted and transformed states;
• categorical separation from every foreground role in both states, plus sampled-corner floors for deep-profile category spacing, foreground clearance, and background contrast;
• terminal day / night capacities 6 / 6, 6 / 6, 4 / 4, 3 / 3;
• no more than 0.15 ΔEOK between each authored transformed accent target and the transformed serialized color that reproduces it;
• at least 4.5:1 transformed contrast for foreground-capable ANSI slots against the terminal base background (bg_0);
• transformed contrast floors of 4.5:1, 3.5:1, and 2.4:1 for fg_0, fg_1, and fg_2 respectively on every background; fg_1 is limited to larger supporting text or graphics, and fg_2 to nonessential metadata or decoration—not body text;
• profile-specific accent-distance floors against each foreground role in commanded and transformed states, so an accent cannot hide a collision in the supporting or muted tier;
• connected foreground ladders with bounded adjacent distances, balanced adjacent lightness gaps, lightness-dominant steps, aligned chroma vectors, mode-aware chroma direction within a quantization tolerance, and narrow commanded/transformed hue spans;
• dark-mode commanded relative-luminance caps of 0.003, 0.005, 0.009, 0.013, 0.020, and 0.021 across the six-step ladder;
• at least 1.8 ΔEOK between adjacent transformed dark-surface ladder steps and 2.8 ΔEOK between adjacent transformed light-surface steps;
• transformed primary-text floors of 6.8:1, 5.65:1, and 5.3:1 across every surface in the 3400 K, 2000 K, and 1200 K dark families, plus 5.0:1 for 3400 K Light;
• at least 6.0 ΔEOK across each transformed background ladder from bg_0 to bg_5, tightened to 15.0 ΔEOK for 3400 K Light;
• 256 unique float samples per sequential map, with monotonic lightness in both display states, transformed step CV no greater than 0.0001, transformed max:min step ratio no greater than 1.001, and the deep commanded CV tightened to 0.11 at 2000 K and 0.15 at 1200 K;
• exact recomputation of the four ±5% green/blue sensitivity corners; and
• exact regeneration of JSON, CSS, themes, diagrams, specimens, and diagnostics.

References

• Gruvbox, original project and palette: warm neutrals, sufficient contrast, and restrained accents for sustained use.
• MIT 6.813, Color: chromatic aberration, the lower spatial resolution of chromatic channels, sparse color, and restrained saturation.
• Fan et al. (2024), The Effect of Ambient Illumination and Text Color on Visual Fatigue under Negative Polarity: controlled evidence that text color and ambient illumination affect fatigue under dark polarity.
• EU Data Visualisation Guide, Colour for categories: limit categorical count and avoid overly saturated or bright colors.
• Datawrapper, What to consider when choosing colors: use neutral structure, protect small-text contrast, and add redundant cues.
• Redshift, pinned color-ramp implementation and temperature table.
• Björn Ottosson, A perceptual color space for image processing.
• Matplotlib, Choosing Colormaps.
• W3C WAI, Understanding contrast minimum.

License

MIT © 2026 Michael Woods.