DitherMe | GlitchNodes

If you've ever wanted to argue about whether Atkinson or Burkes is the superior error-diffusion kernel, DitherMe is your node. It implements over thirty dithering algorithms in one place - the full error-diffusion family, Bayer ordered dithering, clustered and dispersed dots, halftone lines and dots, blue noise, and a few genuinely weird ones like Hilbert curve and spiral dithering. It's the deepest node in the GlitchNodes pack, and the one that turns a boring render into something that looks like a screen-printed poster.

Under the hood it's honest image-processing: error diffusion (Floyd-Steinberg, Jarvis-Judice-Ninke, Stucki, Atkinson, Burkes, Sierra and its variants) spreads quantization error to neighboring pixels for that organic "printed" look, while the ordered/Bayer paths use fixed matrices for clean, mechanical patterns. What makes it more than a novelty is color_mode: monochrome, duotone, tritone, or indexed. The duo/tritone modes with the optional shadow_color / midtone_color / highlight_color (plus per-region brightness sliders) give you full control over the two- or three-color look - that's how you get a properly stylish risograph effect instead of default black-and-white.

The inputs that matter

• algorithm - the big dropdown. Start with floyd_steinberg; graduate to stucki or atkinson when you want a finer texture. hilbert_curve and spiral are worth one try each just to see what they do.
• color_mode - monochrome / duotone / tritone / indexed. This changes the whole character of the output; indexed uses palette_colors (2–256).
• effect_size (0.1–10) - scales the dither pattern density. Higher = chunkier, more visible patterns.
• threshold (0–1) - base threshold for dither decisions; the safety valve for brightness.
• sharpen / blur / noise - pre-processing applied before dithering. A little sharpen makes fine detail survive; a little noise gives film grain.
• temporal_coherence (0–1) - frame-to-frame blending for video. This is the pro move: without it, dithered animation shimmers like a bad GIF; a bit of coherence keeps it stable.
• seed (optional) - reproducible noise for the random/blue-noise paths.

Output is image (IMAGE) → PreviewImage/SaveImage. Batches are supported, and the temporal controls are clearly aimed at video work.

Installing it

Part of the GlitchNodes pack - install once, get all 19 nodes. ComfyUI Manager → Install Custom Nodes → search "GlitchNodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/pxl-pshr/GlitchNodes
# restart ComfyUI

Dependencies: numpy, torch, Pillow, opencv-python, scipy. No model downloads - this one is pure CPU-friendly NumPy.

Gotchas

The sheer number of algorithms is the trap: people pick the wildest one (spiral, wave_interference) and wonder why their render is unrecognizable. Treat it like a toolset - error diffusion for portraits, ordered/Bayer for technical art, halftone for posters. Also, some of the special-effect algorithms are slow on large images, so drop effect_size before you run it on a 4K batch. And remember the pack's standing disclaimer: "highly experimental and may break or even error" - if a particular algorithm chokes, try a smaller image or a sibling algorithm.