post op• Data kinds: image2d → image2d
• Call: import gfx2d; gfx2d.dither(img, levels=2, method='ordered', matrix_size=4) (or opsgfx2d.get("dither"))
Quantise to `levels` values per channel while preserving the local mean.
`method`:
• `"ordered" — Bayer threshold matrix of side matrix_size` (a power of
two, 2..16). The mean error has a closed-form bound: for a uniform
patch the output mean is within `0.5 / (matrix_size**2 * (levels-1))` of
the input, because the fraction of thresholds crossed is the input
fraction rounded to the nearest `1/matrix_size**2`. At the defaults
(4, 2 levels) the bound is 0.03125; the suite checks it holds for 101 grey
levels x 3 matrix sizes, with zero violations.
• `"floyd_steinberg"` — error diffusion with the 7/3/5/1 sixteenths of
Floyd & Steinberg 1976, serial and directional.
Which method preserves the mean better depends on the image, and this
docstring will not pretend otherwise. Measured on a 128-step horizontal
ramp: ordered is exact (0.0) at 2, 3 and 4 levels and drifts to 1.1e-3
at 8 levels, while Floyd-Steinberg is 7.3e-4 at 2 levels and improves to
4.1e-5 at 16. Ordered wins where the level lattice divides the ramp, error
diffusion wins where it does not.
Accepts `(H, W) or (H, W, 3|4)`; an alpha channel is quantised too,
because a dithered sprite with a smooth alpha is exactly the case that
motivates this.
Output values lie on the lattice `k / (levels - 1)`.
• Sample-data catalog (download URLs / licences) — 2-D uses skimage.data (BSD/public domain) plus synthetic images; 3-D lists download URLs for real data sources (Stanford, PDS, …).
• Operator provenance and references — the sources of the research/methods this op family came from.
• The canonical algorithm (author, year) and its uses are named in the family usage guide above.
• gfx2d_scene — py -3.11 examples/gfx2d_scene.py
image2d as input)post)bloom · vignette · chromatic_aberration · film_grain · color_lut · color_grade · palette_quantize
*Provenance: gfx2d.py — GFX2D operator registry. This per-op note is generated by tools/opdocs.py md (do not hand-edit).*
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.