Audio Feature Extractor ⚡🅡🅞🅣🅘

This is the node that makes "audio-reactive" actually happen in this pack. It analyzes a waveform and outputs a per-frame FEATURE signal - a curve any Flex target (FlexImageWarp, FlexImageDepthWarp, FlexImageParallax, and others across the pack) can plug into its opt_feature input and use to drive one of its own parameters. Without this node (or a source node like it), the Flex system has nothing to react to.

The five extraction methods

extraction_method is the whole personality of the node, and the author's own tooltip descriptions are worth quoting directly because they tell you exactly when to reach for each one:

• amplitude_envelope - overall loudness changes, great for syncing with dramatic moments.
• rms_energy - continuous energy level, smoother than amplitude, good for sustained effects.
• spectral_centroid - brightness of the sound, high for sharp/crisp sounds, low for bass/warm sounds.
• onset_strength - detects new sounds/beats, perfect for rhythmic effects.
• chroma_features - musical note content, useful for harmony-based effects.

If you're not sure where to start: onset_strength for anything you want to look punchy and on-beat, rms_energy for anything you want to feel smooth and breathing rather than jittery. amplitude_envelope and rms_energy sound similar in description but behave differently in practice - amplitude tends to read jittery on a busy mix, RMS is noticeably smoother because it's an energy average rather than an instant reading.

Inputs and outputs that matter

• audio and extraction_method - the source and the analysis method.
• frame_rate - the video rate you're targeting.
• frame_count (default 0) - leave at 0 and it auto-calculates frame count from the audio's length and your frame rate; set it explicitly and the extracted feature gets interpolated to match that target length instead.
• width/height - sizes the internal FEATURE tensor; match your pipeline's resolution.

Outputs: feature (FEATURE, wire into any Flex target's opt_feature) and frame_count (INT, handy to reuse elsewhere so your batch length stays consistent without recomputing it).

Installing it

ComfyUI Manager: search RyanOnTheInside, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/ryanontheinside/ComfyUI_RyanOnTheInside
cd ComfyUI_RyanOnTheInside
pip install -r requirements.txt

No model download needed - this is signal analysis, not a learned model.

Where people get burned

Picking an extraction_method is genuinely trial-and-error the first few times - there's no wrong answer, but a method that doesn't match your intent (say, chroma_features on a percussion-heavy track with little melodic content) will produce a feature curve that doesn't do much. If your downstream effect looks too tame or too wild once you've got a signal wired in, that's not this node - it's strength, feature_threshold, and feature_mode on the target node. This node's only job is producing the raw curve; how hard that curve gets applied is entirely the target's decision.