🎵📊 Audio Visualizer

The Audio Visualizer (🎵📊) is the weird one in the pack, and I mean that as a compliment. There's no checkpoint, no sampler, no VRAM budget. It reads a waveform and draws it - particles, bars, a circular waveform - into an image you can do whatever you want with. If you've ever wanted a music video where the visuals come from the track rather than being generated by another model, this is a cheap way to get there.

Worth knowing where it sits in 2026's audio world: the fancy route is a native audio+video model like LTX-2, which invents picture and sound in one pass. This node is the opposite philosophy - deterministic drawing driven by your actual audio, which means the output is exactly as long as your song and doesn't drift or hallucinate.

How it works

It's numpy and Pillow, not diffusion. The node takes the audio tensor and sample rate, computes total_frames = duration × framerate, then calls a visualizer script once per frame. Each script is a plain .py file in nodes/audio_visualizers/ that implements a visualize() function and gets the audio tensor as (batch, channels, samples).

Two things about that loop shape the whole node:

• The scripts are stateful. Particles and waveforms carry position between frames - that's what makes the motion look like motion. Loaded fresh, everything starts at frame zero. That's why seed exists at all: the node isn't using it as a random seed, and setting it to 0 reloads every visualizer module, wiping the global state for a clean animation start.
• Only one output is real per run. Set output_type and the matching output carries data; the other two return nothing.

Inputs that matter

• output_type - image (one frame), image_batch (every frame as a batch), or video (a VIDEO object).
• visualizer_script - whatever .py files are sitting in the visualizers folder. Out of the box you get bouncing bars, particles, umbrella, waveform circle and waveform line.
• scale - sensitivity. Crank it for quiet tracks; higher makes quiet passages move more.
• stereo_to_mono - mean, left, or right. If your track is stereo, it gets collapsed before drawing, so pick which side you actually want for a lopsided mix.
• seed - with output_type: image this is your frame selector: the node takes seed % total_frames and renders that frame. Set it to 0, switch Control After Generate to increment, and set your batch count to the number of frames you want - you'll walk through the timeline one frame per queue.

framerate, width and height (steps of 64, up to 8192) round it out. Width and height are your render resolution, not a suggestion - a 4096-wide batch of a four-minute track is a lot of pixels.

Wiring the outputs

image goes into anything that takes an image: an img2img pass, a control image, a compositing node. image_batch is for video nodes - VHS Video Combine takes the frames and the audio, so check the framerate matches or your audio will slide out of sync. video is a VIDEO object for ComfyUI's native Save Video node, which is the least fiddly of the three.

Install

Search ComfyUI-mnemic-nodes in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/MNeMoNiCuZ/ComfyUI-mnemic-nodes

Restart. No models to download for this node. The pack's requirements.txt does pull imageio (used to encode the video output, mp4/libx264 by default), plus opencv-python, transformers, tiktoken and friends - that's the pack-wide install cost, not something this node specifically needs.

Where people get burned

The batch mode warning is not a joke. image_batch on a full track is thousands of images in one tensor. Output it to a Preview Image node and you'll watch ComfyUI chew through memory until it stutters or dies. Test with image_batch on a ten-second clip first, or use image mode and step frames.

Umbrella is much slower than the rest. That's straight from the author, and it holds - the circular bar drawing does far more work per frame than a waveform line.

Custom scripts fail silently-ish. A broken script throws inside the loop and the node substitutes a black frame rather than stopping, so you can render a whole video of black and only notice afterwards. Watch the console.

If you do write your own, the contract is simple:

def visualize(audio_data, frame_number, framerate, width, height, scale):
    # return a float32 numpy array shaped (height, width, 3), values 0.0-1.0

Drop it in the audio_visualizers folder, restart, and it shows up in the dropdown.