Draw Audio Signal

ARDrawSignal (display name "Draw Audio Signal") is the eyeball node of the ComfyUI-AudioReactive pack. Feed it any SIGNAL and it renders a matplotlib plot of that signal and hands you back a normal ComfyUI IMAGE you can preview or save. It exists because audio features are invisible - you can't tell whether your RMS envelope or onset-strength curve actually has the shape you think it does until you look at it.

This pack is dmarx's port of his old video-killed-the-radio-star notebook, and in that notebook "display the signal" was the step you did constantly while tuning a music-video pipeline. You'd extract a feature, squint at the plot, adjust, repeat. This node is that feedback loop, moved into the graph. It's a debug/verification tool, not a destination - you use it to check what your feature chain is producing before you wire it into SignalToCurve (the node that turns a SIGNAL into a KEYFRAMED_CURVE for driving animation).

How it works

The node calls the pack's draw_signal(y, sr, raw=...) helper. The raw flag comes straight from the signal dict - a SIGNAL carries is_raw=True when it's a raw waveform (as emitted by ARReadAudio) and False for feature-level signals. The flag only changes the x-axis:

• raw → librosa.samples_to_time(...) - x-axis is seconds, sample-accurate.
• not raw → librosa.frames_to_time(...) - x-axis assumes you're plotting frame-based features.

Then it plots with matplotlib, saves the figure to a PNG in memory, converts it to a torch tensor, and returns it in ComfyUI's [B, H, W, 3] layout so PreviewImage/SaveImage just work. No file output, no interactive widget - it's a plot-as-image.

Inputs and outputs

• signal (SIGNAL, marked forceInput) - any signal from the pack: ARReadAudio's raw waveform, or the output of any Op* operator.
• IMAGE - wire to a PreviewImage node to see it inline, or SaveImage to keep it.

Installing it

Same as every node in the pack - Manager (search "AudioReactive") or:

cd ComfyUI/custom_nodes
git clone https://github.com/dmarx/ComfyUI-AudioReactive

then restart. First start is slow: the pack auto-installs scipy, scikit-learn, librosa, loguru, and librosa pulls in numba - budget a couple of minutes.

Common issues

• Blank or missing x-axis meaning. If you plot a frame-based feature - Rms, Novelty, Predominant_Pulse all emit per-frame curves - the pack keeps is_raw=True on them, so this node labels the axis as if each value were a sample. The shape of the curve is right; the time scale may be way off (a 10-second track's ~430 RMS frames will render as if it were a fraction of a second). Treat the plot as "what does the curve look like," not "what's the exact timestamp." Annoying, but it won't hurt your downstream animation.
• Everything downstream explodes. If you feed this node the output of one of the pack's parameterized-but-unwired operators (Bandpass, Smooth, Sustain, Modulo, Quantize, Pow, Clamp), its y is actually a Python function object, and matplotlib will throw a TypeError. That's a known half-finished bit of the pack - see those nodes' pages.
• The pack won't load - nodes/__init__.py imports keyframed at startup and it isn't auto-installed. ModuleNotFoundError: No module named 'keyframed' → pip install keyframed (or install ComfyUI-Keyframed via Manager).

Where people get burned: they treat Draw Audio Signal as an output node that finishes the pipeline. It doesn't. It's a scope for your feature engineering. Use it to confirm the curve you're about to hand to SignalToCurve is spiky in the right places, then get it out of the way.