kubakub shader

Here's the thing nobody tells you when you start a projection-mapping project: half of what ends up on the wall isn't AI at all. Patterns, light, masks, grids, a band running over the windows, tape artefacts, a wire mesh of the building. You can generate those with a diffusion model - people do, at a minute a frame, different every time, which for a looping install is a nuisance. Or you can draw them with a shader: instant, deterministic, loops perfectly, no model to load.

kubakub shader is the second option, built into the pack. Pick one of 74 presets sorted by type, adjust it with sliders, or paste a shader from shadertoy.com and use that. It gives you one picture per frame on images, the same as a mask on masks, and - when the preset is a 3D scene - a depth pass on depth.

How it runs, and why there's nothing to install

These are GLSL ES 3.00 fragment shaders executed on ComfyUI's own GL context - the same ANGLE-based one that ComfyUI's core GLSL node uses (comfy_extras/nodes_glsl.py). That's the whole reason a pack with a GPU shader engine adds exactly one package to your install (OpenCV), and why the node can say "no AI model" honestly. The context is the pack's own, because core's lives on the thread that ran it, and GL is imported on the first render only.

Each frame gets the usual uniforms plus the pack's own: iTime driven by speed, kLoop (a 0→1 phase, run loops times across the clip) which is what the presets actually animate on, the sound as kLevel, kLow, kMid, kHigh, kBeat and kBeats, the field and mask, your regions, and up to four pictures as iChannel0–iChannel3. If you paste code, anything declared as uniform float name; // default min max becomes a control - you set values in the controls field as scale = 5, warp = 2 and the report tells you what names and ranges exist. You don't have to read the shader to use it.

The presets are plain .glsl files in the repo, sorted by type: patterns, raymarched 3D scenes (melting forms, a colonnade, the facade as a relief), looks that work on your regions (each window its own colour, pace or pattern), picture looks, screen and tape (VHS, composite video, 8-bit), grade and lens, picture-with-depth effects, and overlays that run over time. There's a button under preset that shows them as moving pictures, which is the only sane way to shop in a list of 74.

Inputs you'll actually set

preset and code are alternatives: the preset is used while code is empty, so pasting code overrides it. controls is where you dial the preset in. width and height at 0 mean "the size of the first connected picture or mask, else 1280×720". fps, seconds and speed define the clip - note the split: presets run on loops, speed moves iTime, and setting speed to 0 gives you a still.

Then the interesting connections. image0 is your picture (a batch plays as a clip, one per frame): presets designed to work on a picture redraw it, everything else gets laid over it according to on_picture and picture_amount. mask is where the shader shows - kMask(uv) in the shader, and with clip on it also cuts the pictures black outside. field is a grey ramp from kubakub mask field, the path a band runs along. audio is what it hears, and connecting a sound makes a preset with a sound or beat control jump to 1 unless you type a value. regions with only is how the "on your regions" presets get their windows. mask_from decides whether masks is the shader's alpha or its brightness - worth remembering, since a pasted Shadertoy has no alpha (it writes none on shadertoy.com either), so brightness it is.

Outputs: images, masks, fps, audio (cut to the frames), frame_count, depth (from the 3D presets, near = white; one black frame from shaders that write none), and report.

Install and paste

Pack install is the usual:

cd ComfyUI/custom_nodes
git clone https://github.com/kubakubkub/ComfyUI_KubaNodes
python -m pip install -r ComfyUI_KubaNodes/requirements.txt

Restart; ComfyUI 0.37.0+ required, since the pack is written against the newer V3 node API - the comfyui-ecosystem doc explains that, and it's why you'll find no NODE_CLASS_MAPPINGS in a modern pack like this one. If ComfyUI's own GLSL node doesn't run on your machine, this node won't either; that's the shared dependency. The workflow shader.json runs with nothing connected.

Pasting a Shadertoy: open it, click into the code, Ctrl+A, Ctrl+C, paste into code, queue. If there's a Common tab, paste that first and Image below it. Buffer-tab shaders (Buffer A–D) work too if you paste each tab under a line naming it, and the node assumes the usual iChannel0 = Buffer A wiring unless you say otherwise on that line. It won't download links, and it says so.

Where it bites

The node is honest about what can't work: cubemaps, 3D textures, sound shaders, keyboard, webcam, and iMouse (which reads as never pressed). It also refuses the surprises rather than throwing a graphics error. What it can't fix is RAM: 250 frames at 3840×2160 is about 33 GB, and the node will refuse a clip that doesn't fit and tell you what it needs - 3D presets take roughly twice that. On a 5070 Ti a simple preset is around 14 ms a frame at 1080p, warped noise about 32 ms, and a raymarched 3D preset around 100 ms at 1080p, so the shader is rarely the bottleneck; copying frames in and out is. And keep loops in mind: if your clip seems to move at a weird pace, that's the control doing the moving, not speed.