Pixel Art Studio

There's a well-earned community consensus on AI pixel art, and it's that diffusion models don't produce it - they produce something that looks like it from across the room: off-grid, blurry, hundreds of colors, inconsistent pixel sizes. Real pixel art is a deterministic post-process: grid-aware downscale, palette quantization, then clean upscaling. Pixel Art Studio is that entire pipeline in one node, and it's built for the case the consensus usually ignores - doing it to a video sequence without the result boiling and flickering frame to frame.

If you've ever rendered something in a pixel-art style and zoomed in to find your "pixels" aren't uniform or your palette is 300 colors, this is the fix. Feed it frames, get back clean, crisp, limited-palette pixel art.

How it works

It's three stages fused, all with temporal stability bolted on:

1. Block reduction. Each block of pixel_size×pixel_size pixels is collapsed to one color via a per-block winner-take-all vote over 32 color bins - dominant-color voting, not averaging, which is exactly what the KB's pixel-art doctrine says or you get mush. kernel_size (default 5, equal to pixel_size) controls the sampling window: equal means exact tiling, larger overlaps neighboring blocks for a softer, chunkier look.
2. Global palette. A k-means palette is fit once across all frames on the reduced blocks (256/128/64/48 colors), so the palette can't flicker frame to frame. Every frame's art pixels snap to the nearest palette entry in Oklab, with per-pixel temporal hysteresis so a static pixel doesn't pop a pixel thicker/thinner when its color drifts across a palette boundary.
3. Output sizing. Optional width/height (0 = input size), resized with nearest-to-integer-multiple + box down so the pixel grid stays uniform at fractional scales.

The video-first thinking shows everywhere: stability (0–1) is temporal hysteresis on the block vote - a block keeps its previous winner unless clearly beaten. And loop does a hidden warm-up pass so looping footage doesn't snap at the last-to-first wrap. Turn loop off for non-looping video.

Inputs that matter

• frames (IMAGE) - can be a single image or a video batch.
• pixel_size (default 5) - the chunky-ness. Higher = chunkier.
• num_colors (256/128/64/48) - the palette budget. 256 keeps it painterly; 48 reads as properly retro.
• edge_style - hard (default) = crisp sprite edges with binary alpha, matching hand-drawn assets; soft = area-averaged, antialiased look. For sprites, hard.
• stability (default 1) - video smoothness. 0 disables hysteresis.
• loop (default true) - seamless loops; disable for non-looping footage.
• alpha_threshold (default 0.58) - for alpha input; where "50% opaque" sits in the graded coverage.

Output is pixel_art (IMAGE), same batch as input. Frames with an embedded alpha channel come out with one.

Installing it

Part of Link Comfy Nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/Mister-Link/link-comfy-nodes
cd link-comfy-nodes
pip install -r requirements.txt

or ComfyUI Manager → "Link Comfy Nodes", restart. The palette stage needs scikit-learn (it's in requirements.txt), which you won't have from stock ComfyUI - that's the most common install stumble.

Common issues

• Import error about sklearn - install the pack's requirements; scikit-learn doesn't come with ComfyUI.
• Stills look fine but video shimmers - raise stability toward 1. The hysteresis is what kills the boil.
• Looping footage has a visible seam - loop defaults to true for a reason; if you turned it off, turn it back on. Conversely, non-looping footage with loop on gets a warm-up pass that wastes a little time for no benefit.
• Pixels aren't uniform at fractional output sizes - the node handles it internally, but if you set a width/height that isn't a clean multiple, expect the grid to be nearest-integer rather than perfectly square.

One thing it won't do is invent a period-correct palette like PICO-8's 16 colors - you get k-means at your chosen budget, not a named hardware palette. For everything else - grid, palette, stability - this is the single node that replaces the three-step dance most people string together manually.