A program for reducing images to small palettes.

The primary innovation is the introduction of a new ordered dithering algorithm, called simplex dithering, that works with arbitrary palettes. Unlike many other such algorithms, it is provably "correct": as area increases, the average color of the resulting dithered image will converge to the average color of the original image.

This accuracy is the primary goal of this implementation. As a result, quality sometimes suffers compared to other palletization software. To remain accurate, the colors used don't just need to resemble the target colors, but they must be in some sense more extreme than the target colors so that they can average out to the correct middle value. Including the corners of the RGB cube (white, red, green, blue, yellow, cyan, magenta, and black: this is also the 3bit palette selectable in the program) is always sufficient to approximate anything, but isn't always pretty. Traditional palette selection, using K-Means or related methods, aims to produce a set of colors that are in the middle of the targets, minimizing distance and improving the quality of simple quantization, but it fails to represent the full range of colors. This trades high frequency noise (which is more noticeable to the human eye) for low frequency error.

Installation

If you already have cargo installed, you can use it directly to install dither:

cargo install --git https://github.com/gereeter/dither

Usage

USAGE:
    dither [OPTIONS] <IMAGE>

ARGS:
    <IMAGE>    Sets the image to dither

OPTIONS:
    -a, --algorithm <ALGORITHM>    Chooses the dithering algorithm to use [default: simplex]
    -b, --bias <BIAS>              Chooses the bias pattern for ordered dithering algorithms
                                   [default: plastic+triangle]
    -c, --colors <PALETTE_SIZE>    How many colors to use in a procedural palette [default: 16]
    -d, --distance <DISTANCE>      Chooses how to calculate how far apart colors are [default:
                                   CIEDE2000]
    -h, --help                     Print help information
    -o, --output <OUTPUT>          Sets where to write the dithered file to [default: out.png]
    -p, --palette <PALETTE>        Chooses the palette to quantize to [default: simplex]
    -s, --summarize                Print a summary of the palette colors used in the output
    -V, --version                  Print version information

For example, if you want to dither an image to only use websafe colors, you might run

dither --palette websafe --output image_dithered.png image.png

Available algorithms include

• nearest, which implements basic image quantization mapping every pixel to its nearest palette color, often producing large blocks of solid color,
• simplex, the new innovation of this program, which mixes up to 4 palette colors to perfectly accurately approximate an image with an ordered / stable dither pattern, and
• floyd-steinberg, implementing Floyd-Steinberg dithering, an error diffusion dithering algorithm, which can produce excellent results but can vary wildly with small changes to the input, and can sometimes result in "worm" artifacts where a specific error is pushed all into a line.

Bias patterns control the look and feel of the resulting dithered image, and only apply to ordered dither algorithms. (Most algorithms implemented are ordered dithers, but, e.g. Floyd-Steinberg dithering is unaffected.)

• The default, plastic+triangle, produces a very even fabric-like pattern, based on the suggestion in The Unreasonable Effectiveness of Quasirandom Sequences.
• For an old-school look, use bayer256, which implements the classic Bayer matrix ordered dithering on a large scale. Smaller Bayer matrices are available, but they can reduce quality for little gain.
• For something approximating halftoning, use dot8, which deliberately groups colors together into larger artifacts. This was mostly intended for debugging, and so its quality isn't ideal for producing that look.
• For a mottled look, consider using random, which uses actual pseudorandomness instead of quasirandomness to select between palette colors. It is very hard to control, but running it multiple times can give several options to choose between.

For manually editing the results or embedding in other formats, the --summarize option can give insight into the results of dithering. For example, to get a distribution of colors useful for representing a single color, using a solid color image as input with --summarize will print out exactly the mix of palette colors necessary to produce that color. (There may be a dedicated function for this added later.)