dithr 0.3.0

Buffer-first rust dithering and halftoning library.
Documentation

dithr

crates.io

Original Dithered
Before dithering After dithering

Before (left) and after (right) using yliluoma_2_in_place.

Buffer-first rust dithering and halftoning library.

Quantizing grayscale/RGB/RGBA buffers without dithering creates visible banding and contouring. dithr provides deterministic ordered dithering, diffusion, stochastic binary methods, palette-constrained workflows, and advanced halftoning methods over typed mutable slices.

Overview

  • Buffer-first API: Works directly on mutable pixel slices with explicit width, height, and stride.
  • Typed formats: Supports u8, u16, and f32 sample types across Gray, Rgb, and Rgba layouts.
  • Quantization control: Uses QuantizeMode for grayscale levels, RGB levels, palette mapping, or single-color workflows.
  • Broad algorithm coverage: Includes stochastic, ordered, palette-oriented ordered, diffusion, variable diffusion, and advanced halftoning families.
  • Palette workflow: Includes Palette<S> and IndexedImage<S> for constrained output and indexed results.
  • Optional integrations: image adapters for DynamicImage workflows and rayon parallel wrappers for selected families.

Installation

cargo add dithr

[dependencies]
dithr = "0.3.0"

cargo add dithr --features image
cargo add dithr --features rayon

Quick Start

use dithr::{gray_u8, QuantizeMode, Result};
use dithr::ordered::bayer_8x8_in_place;

fn main() -> Result<()> {
    let width = 64_usize;
    let height = 64_usize;
    let mut data = Vec::with_capacity(width * height);

    for y in 0..height {
        for x in 0..width {
            let value = ((x + y * width) * 255 / (width * height - 1)) as u8;
            data.push(value);
        }
    }

    let mut buffer = gray_u8(&mut data, width, height, width)?;
    bayer_8x8_in_place(&mut buffer, QuantizeMode::gray_bits(1)?)?;

    assert!(data.iter().all(|&v| v == 0 || v == 255));
    Ok(())
}

Core Data Model

dithr is organized around a small set of types that are shared across algorithm families.

  • Buffer<'a, S, L>: Mutable typed view of image data (S = sample type, L = layout).
  • BufferKind / PixelFormat: Runtime format metadata (PixelFormat is an alias of BufferKind).
  • Palette<S>: Palette storage for fixed-color workflows (1 to 256 entries).
  • IndexedImage<S>: Indexed output (Vec<u8> indices) paired with a typed palette.
  • QuantizeMode<'a, S>: Common quantization target model used by ordered/diffusion/stochastic/dot/Riemersma families.
  • Error / Result<T>: Crate-level error and result surface.

Typed Buffers, Sample Types, and Layouts

Supported runtime kinds:

  • Gray8, Rgb8, Rgba8
  • Gray16, Rgb16, Rgba16
  • Gray32F, Rgb32F, Rgba32F

Typed buffer aliases:

  • GrayBuffer8, RgbBuffer8, RgbaBuffer8
  • GrayBuffer16, RgbBuffer16, RgbaBuffer16
  • GrayBuffer32F, RgbBuffer32F, RgbaBuffer32F

Constructor helpers:

  • Gray: gray_u8, gray_u16, gray_32f
  • Rgb: rgb_u8, rgb_u16, rgb_32f
  • Rgba: rgba_u8, rgba_u16, rgba_32f
  • Packed variants: gray_u8_packed, rgb_u16_packed, rgba_32f_packed, etc.

Generic constructors remain available on Buffer:

  • Buffer::new_typed(...)
  • Buffer::new_packed_typed(...)
  • Compatibility constructors with runtime kind checking:
    • Buffer::new(...)
    • Buffer::new_packed(...)

Choosing a Quantization Mode

QuantizeMode<'a, S> is the canonical quantization model:

  • GrayLevels(u16)
  • RgbLevels(u16)
  • Palette(&Palette<S>)
  • SingleColor { fg: [S; 3], levels: u16 }

Convenience constructors:

  • Generic:
    • QuantizeMode::gray_levels(levels)
    • QuantizeMode::rgb_levels(levels)
    • QuantizeMode::palette(&palette)
    • QuantizeMode::single_color(fg, levels)
  • u8 compatibility helpers:
    • QuantizeMode::gray_bits(bits)
    • QuantizeMode::rgb_bits(bits)
  • Shared conversion helper:
    • levels_from_bits(bits)

Use GrayLevels/gray_bits when output should be grayscale levels, RgbLevels/rgb_bits for uniform per-channel color quantization, Palette for strict membership in a fixed color set, and SingleColor for foreground-scaled tonal output.

Dithering and Halftoning Method Families

Binary stochastic

Fast binary dithering with fixed or randomized threshold behavior.

  • threshold_binary_in_place
  • random_binary_in_place

Parallel variants (rayon feature):

  • threshold_binary_in_place_par
  • random_binary_in_place_par

Ordered methods

Deterministic threshold-map methods with predictable structure and straightforward benchmarking.

Bayer:

  • bayer_2x2_in_place
  • bayer_4x4_in_place
  • bayer_8x8_in_place
  • bayer_16x16_in_place

Cluster-dot:

  • cluster_dot_4x4_in_place
  • cluster_dot_8x8_in_place
  • void_and_cluster_in_place

Custom map:

  • custom_ordered_in_place
  • adaptive_ordered_dither_in_place
  • space_filling_curve_ordered_dither_in_place
  • ranked_dither_in_place
  • image_based_dither_screen_in_place
  • polyomino_ordered_dither_in_place
  • stochastic_clustered_dot_in_place
  • am_fm_hybrid_halftoning_in_place
  • clustered_am_fm_halftoning_in_place
  • blue_noise_multitone_dither_in_place

Parallel variants (rayon feature):

  • bayer_2x2_in_place_par
  • bayer_4x4_in_place_par
  • bayer_8x8_in_place_par
  • bayer_16x16_in_place_par
  • cluster_dot_4x4_in_place_par
  • cluster_dot_8x8_in_place_par
  • custom_ordered_in_place_par

Palette-oriented ordered (Yliluoma)

Ordered methods designed for fixed-palette workflows.

  • yliluoma_1_in_place
  • yliluoma_2_in_place
  • yliluoma_3_in_place

Classic diffusion

Scanline error diffusion kernels for higher local tonal quality.

  • floyd_steinberg_in_place
  • false_floyd_steinberg_in_place
  • jarvis_judice_ninke_in_place
  • stucki_in_place
  • burkes_in_place
  • sierra_in_place
  • two_row_sierra_in_place
  • sierra_lite_in_place
  • stevenson_arce_in_place
  • atkinson_in_place

Extended diffusion

Additional diffusion kernels with different spread patterns.

  • fan_in_place
  • shiau_fan_in_place
  • shiau_fan_2_in_place
  • block_error_diffusion_in_place

Scope note: block_error_diffusion_in_place is grayscale-only.

Variable diffusion

Tone-dependent coefficient families.

  • ostromoukhov_in_place
  • zhou_fang_in_place
  • hvs_optimized_error_diffusion_in_place
  • gradient_based_error_diffusion_in_place
  • multiscale_error_diffusion_in_place
  • feature_preserving_msed_in_place
  • green_noise_msed_in_place
  • linear_pixel_shuffling_in_place
  • tone_dependent_error_diffusion_in_place
  • structure_aware_error_diffusion_in_place
  • adaptive_vector_error_diffusion_in_place
  • vector_error_diffusion_in_place
  • semivector_error_diffusion_in_place
  • hierarchical_error_diffusion_in_place
  • mbvq_color_error_diffusion_in_place
  • neugebauer_color_error_diffusion_in_place
  • multichannel_green_noise_error_diffusion_in_place

Scope note: variable diffusion methods are grayscale-only except adaptive_vector_error_diffusion_in_place, vector_error_diffusion_in_place, semivector_error_diffusion_in_place, and hierarchical_error_diffusion_in_place, mbvq_color_error_diffusion_in_place, and neugebauer_color_error_diffusion_in_place, and multichannel_green_noise_error_diffusion_in_place, which support Rgb/Rgba with alpha preservation on Rgba.

Advanced halftoning

Specialized methods with narrower scope than the ordered/diffusion baseline.

  • riemersma_in_place
  • knuth_dot_diffusion_in_place
  • optimized_dot_diffusion_in_place
  • direct_binary_search_in_place
  • clustered_dot_direct_multibit_search_in_place
  • direct_pattern_control_in_place
  • hierarchical_colorant_dbs_in_place
  • lattice_boltzmann_in_place
  • electrostatic_halftoning_in_place
  • model_based_med_in_place
  • least_squares_model_based_in_place

Scope notes:

  • direct_binary_search_in_place, clustered_dot_direct_multibit_search_in_place, lattice_boltzmann_in_place, electrostatic_halftoning_in_place, model_based_med_in_place, and least_squares_model_based_in_place require integer grayscale buffers.
  • direct_pattern_control_in_place supports integer Rgb/Rgba buffers; alpha is preserved for Rgba.
  • hierarchical_colorant_dbs_in_place supports integer Rgb buffers.
  • riemersma_in_place, knuth_dot_diffusion_in_place, and optimized_dot_diffusion_in_place support Gray/Rgb/Rgba layouts, with alpha preserved for Rgba paths.

Palette Workflow

dithr keeps palette workflows explicit: define a palette, dither into it, and optionally build indexed output.

use dithr::{rgb_u8, IndexedImage, Palette, Result};
use dithr::ordered::yliluoma_1_in_place;

fn main() -> Result<()> {
    let width = 32_usize;
    let height = 32_usize;
    let mut data = vec![0_u8; width * height * 3];

    for y in 0..height {
        for x in 0..width {
            let i = (y * width + x) * 3;
            data[i] = (x * 255 / (width - 1)) as u8;
            data[i + 1] = (y * 255 / (height - 1)) as u8;
            data[i + 2] = ((x + y) * 255 / (width + height - 2)) as u8;
        }
    }

    let palette = Palette::new(vec![
        [0, 0, 0],
        [255, 255, 255],
        [255, 0, 0],
        [0, 0, 255],
    ])?;

    let mut buffer = rgb_u8(&mut data, width, height, width * 3)?;
    yliluoma_1_in_place(&mut buffer, &palette)?;

    let mut indices = Vec::with_capacity(width * height);
    for px in data.chunks_exact(3) {
        indices.push(palette.nearest_rgb_index([px[0], px[1], px[2]]) as u8);
    }

    let indexed = IndexedImage::new(indices, width, height, palette)?;
    assert_eq!(indexed.len(), width * height);

    Ok(())
}

Built-in palette helpers are also exported:

  • cga_palette()
  • grayscale_2()
  • grayscale_4()
  • grayscale_16()

Optional image Integration

Enable image to adapt image crate buffers into dithr buffers.

Typed adapters:

  • gray8_image_as_buffer
  • rgb8_image_as_buffer
  • rgba8_image_as_buffer
  • gray16_image_as_buffer
  • rgb16_image_as_buffer
  • rgba16_image_as_buffer
  • rgb32f_image_as_buffer
  • rgba32f_image_as_buffer

Dynamic adapter:

  • dynamic_image_as_buffer(&mut image::DynamicImage) -> Result<DynamicImageBuffer<'_>>

Dynamic variants:

  • DynamicImageBuffer::Gray8
  • DynamicImageBuffer::Rgb8
  • DynamicImageBuffer::Rgba8
  • DynamicImageBuffer::Gray16
  • DynamicImageBuffer::Rgb16
  • DynamicImageBuffer::Rgba16
  • DynamicImageBuffer::Rgb32F
  • DynamicImageBuffer::Rgba32F

DynamicImage::ImageLumaA8 and DynamicImage::ImageLumaA16 are promoted to DynamicImageBuffer::Rgba8 and DynamicImageBuffer::Rgba16 during adaptation.

Current manifest configuration enables PNG and JPEG codecs for the optional image dependency.

Optional rayon Integration

Enable rayon for parallel wrappers where available.

Parallelized families:

  • Ordered: all *_in_place_par ordered wrappers
  • Yliluoma: yliluoma_1_in_place_par, yliluoma_2_in_place_par, yliluoma_3_in_place_par
  • Binary stochastic: threshold_binary_in_place_par, random_binary_in_place_par

Current serial-only families:

  • Diffusion (classic/extended/variable)
  • Advanced halftoning

Example Programs

Raw buffer workflows:

cargo run --example gray_buffer
cargo run --example rgb_buffer
cargo run --example indexed_palette

Image workflows (image feature):

cargo run --example image_bayer_png --features image -- input.png output.png
cargo run --example image_palette_png --features image -- input.png output.png

Benchmarks and Development

Bench families (criterion):

  • stochastic
  • ordered
  • yliluoma
  • diffusion
  • advanced
cargo bench --no-run
cargo bench --bench stochastic
cargo bench --bench ordered
cargo bench --bench yliluoma
cargo bench --bench diffusion
cargo bench --bench advanced

Development checks:

cargo fmt --all -- --check
cargo clippy --workspace --all-targets --all-features -- -D warnings
cargo check --all-features
cargo test --workspace --all-features --lib --tests --examples
cargo test --doc --all-features

Limitations and Scope

  • Core processing is in-memory and buffer-first; it does not provide general image editing workflows.
  • Not every algorithm supports every format/layout/sample combination.
  • am_fm_hybrid_halftoning_in_place and clustered_am_fm_halftoning_in_place are grayscale-only.
  • blue_noise_multitone_dither_in_place is grayscale-only.
  • Variable diffusion methods are grayscale-only except adaptive_vector_error_diffusion_in_place, vector_error_diffusion_in_place, semivector_error_diffusion_in_place, mbvq_color_error_diffusion_in_place, and neugebauer_color_error_diffusion_in_place, and multichannel_green_noise_error_diffusion_in_place for Rgb/Rgba.
  • block_error_diffusion_in_place is grayscale-only.
  • direct_binary_search_in_place, clustered_dot_direct_multibit_search_in_place, lattice_boltzmann_in_place, and electrostatic_halftoning_in_place are integer grayscale-only.
  • Parallel wrappers are currently provided for ordered, Yliluoma, and binary stochastic families.

References

License

MIT. See LICENSE.