Djz Pixel Sort V7 Advanced

DjzDatamoshV7 is the pixel sorter you actually want out of the DJZ datamosh family. Where V6 gives you one mode and one direction, V7 gives you four sorting criteria, rotation, a multi-pass mode that stacks them all, and a seed for reproducibility. It's the node to grab when the classic "sort by brightness" look has gotten boring and you want to sort an image by hue so a sunset turns into horizontal color bands, or by edge strength so the whole thing reads like a contour map. Display name in ComfyUI is "Djz Pixel Sort V7 Advanced," and it earns the word.

How it works

Same skeleton as V6 - threshold-based segmentation, then re-sort within segments - but the sort key is selectable. sort_mode picks the property pixels are ordered by:

• luminance - the familiar brightness sort.
• hue - sorts by color angle. This is where things get psychedelic; a gradient of colors reorders itself into smooth rainbow bands.
• saturation - pulls vibrant pixels apart from muted ones.
• laplacian - sorts by edge-detection strength, producing contour-heavy, almost topographic results.

threshold (0–1, default 0.5) controls segment length. The doc's framing is slightly counterintuitive: lower values create fewer, longer sorted segments (more dramatic streaks); higher values make more, shorter segments (granular, subtle). It behaves a bit differently per mode since the threshold applies to whichever value is being sorted, so just treat it as "how broken do I want this."

The controls that change the character

• rotation - rotates the sorting direction in 90° steps, default -90 (vertical sorting). 0 gives horizontal smears; the 180 values flip the direction. The single most useful knob for matching the effect to the image's dominant lines.
• multi_pass - when on, applies all four modes in a fixed order (luminance → hue → saturation → laplacian) instead of just the selected one. Layered and nasty, and noticeably slower.
• seed - makes the run reproducible. Note the sorting itself is deterministic; the seed matters for any randomized variation in the pipeline (and for keeping results stable across ComfyUI re-runs).

Input images, output IMAGE. The node runs in pure numpy/torch - no OpenCV, no ffmpeg, no external binaries - so it's the most portable member of the family. One dependency gotcha: it imports scipy's convolve2d for the laplacian mode, and scipy isn't listed in the pack's requirements.txt. If the node won't load, that's the usual culprit - pip install scipy and restart.

The honest take

For a single static image, a V7 multi-pass or hue sort is one of the best "instant art" effects in the pack - wire an image in, hit queue, and you've got a poster. Over a whole video sequence it can get busy fast, so start with luminance or saturation and only enable multi_pass once you know the shot can take it. And if you compare V7 to the "datamosh" name it carries: it's not datamoshing at all, it's pixel sorting - but that's fine, because it's good at exactly that.