MIDI Feature Extractor ⚡🅡🅞🅣🅘

RyanOnTheInside's whole pitch is "everything-reactivity" - audio, MIDI, motion, proximity, whatever you've got, turned into a signal something else can react to. This is the MIDI half of that: it takes a loaded MIDI file (from this pack's MIDILoader, not a raw audio file - MIDI and audio are different worlds here) and extracts one specific dimension of it into a per-frame FEATURE you can wire into any of the pack's Flex targets.

Why MIDI instead of just analyzing the audio

Audio-reactive effects have to infer things - was that a beat, how loud was it, roughly what pitch. MIDI already knows. It's the actual performance data: exact note timings, exact velocities, exact pitch-wheel and modulation-wheel movements, recorded at the source rather than reverse-engineered from a waveform. If you have the MIDI file behind your music (or you're generating visuals to sync with a MIDI performance directly), this node gets you a cleaner, more precise signal than audio analysis ever could.

How it works

extraction_method is a long dropdown - 16 options - grouped into a few families. Note-level data: Velocity (how hard a note was hit, great for impact effects), Pitch, Note On/Off (precise timing triggers), Note Duration, Note Density (how many notes are active - a proxy for musical complexity). Expression data: Pitchbend, Aftertouch, Poly Pressure. And a full set of MIDI CC channels - Modulation (CC1), Breath (CC2), Foot Controller (CC4), Volume (CC7), Balance (CC8), Pan (CC10), Expression (CC11), Sustain (CC64) - for whenever the source performance was recorded with a controller riding one of those. chord_only narrows extraction to moments where a full chord is present rather than any single note, and notes lets you filter to a comma-separated list of specific MIDI note numbers if you only care about, say, the bassline.

Like the pack's other feature extractors, frame_count at 0 auto-calculates the frame count from the MIDI and your frame_rate; set it explicitly and the extracted feature gets interpolated to that target length instead.

The inputs and outputs that matter

• midi (required, MIDI) - from MIDILoader or another node in the pack that outputs MIDI.
• extraction_method - which dimension of the performance to extract (see above).
• frame_rate (default 30) and frame_count (default 0, auto) - how the note-event timeline maps onto video frames.
• chord_only / notes - narrow which notes count.
• width/height - sizes the internal feature tensor; match your pipeline's resolution.

Outputs: midi (passthrough, so you can keep chaining without re-loading), feature (FEATURE, wire into any Flex target's reactive input, or into FeatureToMask if you just want a plain mask), and frame_count (INT) for reuse downstream.

How to install it

ComfyUI Manager: search "RyanOnTheInside", install, restart. Manually:

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

then restart. MIDI parsing in this pack runs on mido, pinned in requirements.txt - a real, lightweight dependency, not a heavy ML model, so there's nothing to download here.

Common issues & troubleshooting

Feature is flat/empty. The most common cause is picking a CC-based method (say, Sustain (CC64)) for a MIDI file that never actually uses that controller - not every recorded performance touches every CC. Check your source MIDI in a piano-roll editor first if a channel you picked looks suspiciously flat.

Wired in audio instead of MIDI. This node's midi input is strictly the MIDI type - a raw AUDIO waveform won't connect. If you don't have a MIDI file, you want AudioFeatureExtractor or RhythmFeatureExtractor instead, which work directly off audio.

notes filter seems to do nothing. It expects literal MIDI note numbers (comma-separated), not note names like "C4" - double check you're passing numbers, not letters.