Ditherlib
Ditherlib is a Rust library for non-destructive image effects and dithering. Effects can target an entire image or anti-aliased polygon selections, and an ordered pipeline can combine several selected effects into one render.
Source pixels remain immutable after loading. Each render starts from the source and produces a separately owned image, so callers can freely edit and re-render pipelines.

Features
- Greyscale and Gaussian blur
- Threshold, configurable ordered, and deterministic noise dithering
- Print-style halftone screens with six dot shapes and arbitrary palettes
- Independently angled RGB and CMYK process-print screens
- Fourteen error-diffusion presets, from minimal Two-dimensional Knuth through broad Stevenson-Arce
- Custom diffusion kernels, strength, clamping, and scan direction
- Bayer, clustered-dot, line, crosshatch, checkerboard, and dispersed-dot threshold maps
- Built-in 16x16 blue-noise threshold map
- Custom rectangular threshold maps with strength, offset, rotation, and mirroring
- Custom RGB palettes, black-and-white palettes, and monochrome palettes
- Palette derivation from source-image colours with an optional size limit
- RGB, linear RGB, and Oklab palette matching with luminance-only control
- Rectangular logical pixels with grid offsets and five sampling modes
- Whole-image and polygon selections with anti-aliased edges
- Ordered multi-effect pipelines with reusable rendering buffers
- Crate-owned image, result, and error types
Installation
JPEG and PNG support are enabled by default:
[]
= "0.8"
Codec features can be selected individually:
[]
= { = "0.8", = false, = ["png", "webp"] }
Available codec features are avif, bmp, dds, exr, ff, gif, hdr,
ico, jpeg, png, pnm, qoi, tga, tiff, and webp. The processing
algorithms compile without any codec features.
Usage
This example converts the full image to greyscale, then applies a red monochrome ordered dither inside a polygon:
use ;
Pipeline order matters. Each step receives the result of the previous step,
including where polygon selections overlap. Rendering the same pipeline again
always begins from the unchanged SourceImage.
Logical pixel geometry and sampling
Threshold, ordered, noise, and error-diffusion effects share the same logical pixel controls. Width and height use image pixels, and signed grid offsets move the grid right and down for positive values.
use ;
SamplingMode provides coverage-weighted average, centre, darkest, lightest,
and coverage-weighted dominant-colour sampling. Centre sampling uses the
selected source pixel nearest the centre of the image-clipped cell. Darkest and
lightest use RGB luma. Equal dominant-colour coverage resolves to the colour
encountered first. Every mode samples only pixels covered by the active
selection, including anti-aliased polygon intersections and partial cells at
image edges.
Ordered dithering
ThresholdMap provides standard Bayer 2x2, 4x4, and 8x8 maps and validates
custom rectangular maps. Custom values are row-major ranks from zero up to one
less than the map length; repeated ranks are allowed.
Artistic presets are available through ThresholdMap::clustered_dots(),
horizontal_lines(), vertical_lines(), diagonal_lines(), crosshatch(),
checkerboard(), dispersed_dots_3x3(), and dispersed_dots_5x5(). They tile
at the image origin and support the same strength, offset, rotation, and
mirroring controls as Bayer and custom maps.
use ;
Offsets use logical pixels and move the map right and down for positive values. Rotation is clockwise. Mirroring applies horizontally and vertically after rotation. Every transformation remains anchored to the image origin when an effect targets a polygon.
Noise dithering
NoiseDither provides deterministic white-noise and blue-noise threshold
dithering. Its samples are driven by image-origin logical pixel coordinates
and the seed, so rendering a polygon selection does not shift the noise field.
use ;
The fixed map is also available as ThresholdMap::blue_noise_16x16() for
ordered dithering and custom transformations.
Halftone screens
Halftone provides circle, square, diamond, ellipse, line, and cross screens.
Cell dimensions and phase use image pixels, while angles use clockwise radians.
The screen remains anchored to the image origin when used with a selection.
use ;
Run cargo run --example halftone_contact_sheet -- INPUT OUTPUT.png to render
all six shapes. Palette::black_and_white() produces classic monochrome output;
other built-in or custom palettes produce colour screens.
Colour halftoning
ColourHalftone screens colour separations independently. RGB mode thresholds
the red, green, and blue light channels and recombines them directly. CMYK mode
uses under-colour removal to separate cyan, magenta, yellow, and black inks,
screens each ink, then recombines the subtractive channels into RGB output.
use ;
The traditional CMYK preset uses 15° cyan, 75° magenta, 0° yellow, and 45° black screens. The moiré-resistant preset uses 18.4°, 71.6°, 0°, and 45°. Angles are clockwise radians in the API. Offsets use pixels along each rotated screen's axes. Absolute image coordinates and stateless thresholding make the same configuration deterministic across repeated renders and selections.
Run cargo run --example colour_print_comparison -- INPUT OUTPUT.png for a
source, RGB, and CMYK comparison. Tiles retain the source aspect ratio.
Palettes
Palette::black_and_white() provides the familiar two-colour palette.
Palette::monochrome(colour) combines black, the supplied RGB colour, and
white. Eight-level greyscale, Game Boy, CGA, and PICO-8 palettes are also
available through Palette::greyscale(), Palette::game_boy(),
Palette::cga(), and Palette::pico_8().
Colour represents an RGB value with named channels, common colour constants,
and conversions to and from [u8; 3]. Color is an alias for callers using
American spelling. Palette::new accepts either representation:
use ;
Palettes can also be derived from the visible pixels in a source image.
PaletteSize::All retains every distinct colour in first-seen order.
PaletteSize::Limited(n) applies deterministic, frequency-weighted median cut
and chooses representatives that occur in the source. Fully transparent pixels
do not contribute colours.
use ;
Perceptual palette matching
Palettes use gamma-encoded RGB distance by default, preserving the matching and
rendered output from earlier releases. ColourSpace::LinearRgb compares
linear-light channels, while ColourSpace::Oklab compares perceptual lightness
and opponent colour components. Every palette-based effect uses the palette's
configured space.
use ;
let palette = pico_8
.with_colour_space
.with_matching_mode;
assert_eq!;
PaletteMatchMode::Luminance compares only gamma-encoded luma in RGB, physical
relative luminance in linear RGB, or Oklab lightness. Palette order resolves
equal distances.
Error diffusion calculates and distributes errors in the palette's colour space. Independent-channel diffusion is the default. Luminance mode propagates brightness error while leaving chroma error local:
use ;
let effect = new
.with_error_mode;
assert_eq!;
The palette stores converted entries when its colour space is selected. The sRGB transfer constants follow CSS Color 4, and the Oklab matrices follow Björn Ottosson's reference transform.
Error diffusion
ErrorDiffusion accepts a built-in algorithm or a validated custom kernel. It
supports raster or serpentine scanning, diffusion strength from 0.0 through
2.0, and optional per-channel error clamping.
use ;
Custom kernels contain forward-pointing weighted taps and a divisor:
use ;
Built-in presets have distinct grain and edge behaviour:
| Preset | Visual character |
|---|---|
| Floyd-Steinberg | Crisp, balanced detail with a familiar fine grain |
| Atkinson | High contrast with clean highlights, shadows, and clustered dots |
| Jarvis-Judice-Ninke | Soft, finely dispersed grain with smooth tonal changes |
| Stucki | Sharp detail with broad, even error distribution |
| Burkes | Clean two-row texture with less softness than Stucki |
| Sierra | Smooth, balanced grain with gentle transitions |
| Two-Row Sierra | Compact, moderately crisp grain |
| Sierra Lite | Fast, coarse texture with visible directional structure |
| False Floyd-Steinberg | Coarse, strongly directional texture |
| Fan | Compact, left-leaning grain with pronounced diagonal structure |
| Shiau-Fan | Short-tailed texture designed to reduce worm artefacts |
| Shiau-Fan 2 | Longer-tailed grain with smoother highlight and shadow texture |
| Stevenson-Arce | Very fine, dispersed grain with smooth tones and preserved detail |
| Two-dimensional Knuth | Minimal, regular diagonal texture |
Errors
Fallible operations return ditherlib::Result<T>. Use DitherError::kind()
to handle stable ErrorKind categories covering file access, unsupported
formats, decoding, encoding, invalid geometry, dimension mismatches, invalid
parameters, and effect failures. Dependency errors may remain available
through the standard Error::source() method for diagnostics.
JPEG output discards alpha because the format does not support transparency.
Examples
Every example accepts an input path and output path:
The diffusion comparison requires an image with an even width and creates seven
side-by-side comparisons. From comparison-1.png through comparison-7.png,
the left/right pairs are Floyd-Steinberg/Atkinson,
Jarvis-Judice-Ninke/Stucki, Burkes/Sierra, Two-Row Sierra/Sierra Lite, False
Floyd-Steinberg/Fan, Shiau-Fan/Shiau-Fan 2, and
Stevenson-Arce/Two-dimensional Knuth.
The diffusion controls example creates two quadrant comparisons. strengths.png
uses strengths 0.0, 0.5, 1.0, and 1.5 from top-left to bottom-right.
clamps.png uses error limits of 0, 24, 64, and unlimited in the same order.
The ordered comparison creates three quadrant images. maps.png compares Bayer
2x2, 4x4, and 8x8 with a custom 3x2 map. strengths.png compares 0.25, 0.5,
1.0, and 1.5. rotations.png compares 0, 90, 180, and 270 degrees clockwise.
The pattern sheet creates a labelled grid containing one result for each artistic threshold-map preset.
The noise comparison applies white noise to the left half and blue noise to the right half with the same seed and strength.
The palette comparison uses greyscale, Game Boy, CGA, and PICO-8 from top-left to bottom-right. It applies Floyd-Steinberg diffusion with serpentine scanning to every quadrant so the palette is the only variable.
The palette-matching comparison keeps the source in the top-left quadrant and uses RGB, linear RGB, and Oklab matching in the remaining quadrants in reading order. Every processed quadrant uses the same PICO-8 palette.
The pixel-sampling comparison creates five vertical sampling bands and a second image comparing square, wide, tall, and offset-square logical pixels.
Additional examples cover each built-in effect in the examples
directory.