Aeon Music Video

aeon-music-video — Audio-Reactive Music Video Builder

Two complementary scripts for turning a music track + a pile of video clips into a finished music video. Both detect the audio's beats, onsets, RMS, and spectral centroid via librosa, then drive ffmpeg filter chains that pulse brightness, hue, zoom, and chromatic aberration in sync with the music.

When to use which

Prerequisites

• Audio track (FLAC/WAV/MP3) — the song the video reacts to
• A pile of video clips (MP4, any duration) — what gets cut into the final video
• ffmpeg with sendcmd, eq, hue, rgbashift, gblur, showcqt, avectorscope filters (standard build is fine)
• Python 3.10+, librosa, numpy, scipy, soundfile

reactive_compositor.py — general audio-reactive editor

Workflow:

1. Analyze the audio — librosa extracts beat times, onset envelope, RMS, spectral centroid
2. Segment the audio by intensity + brightness (continuous mood map)
3. Match clips to segments by mood bucket (calm / cosmic / building / crystalline / deep / explosive)
4. Sequence clips in order, with crossfades on intensity drops
5. Apply reactive effects:
   • eq filter brightness pulses on every beat (peak +0.25 by default)
   • eq saturation drives by RMS envelope
   • hue rotates by spectral centroid (warmer → cooler colors track frequency)
   • Optional zoom pulses on RMS peaks
6. Optional overlays: showcqt spectrum analyzer + avectorscope stereo image

Usage

# Provide one clip per mood bucket
python scripts/reactive_compositor.py \
    --audio  song.flac \
    --mood-clip calm:cosmic_drift.mp4 \
    --mood-clip building:wireframe_grid.mp4 \
    --mood-clip explosive:fractal_burst.mp4 \
    --mood-clip crystalline:prism_shimmer.mp4 \
    --mood-clip deep:dark_tunnel.mp4 \
    --output  music_video.mp4 \
    --fps 24 --width 832 --height 480

# Or provide a flat list of clips and let the compositor distribute them
python scripts/reactive_compositor.py \
    --audio song.flac \
    --clip clip1.mp4 --clip clip2.mp4 --clip clip3.mp4 --clip clip4.mp4 \
    --output mv.mp4

# Reactive parameters
#   --zoom-peak 0.12        zoom pulse amplitude on RMS peaks (default 0.12)
#   --brightness-peak 0.25  brightness pulse on beats (default 0.25)
#   --hue-max-deg 30        hue rotation range from centroid (default 30°)
#   --no-cqt                disable spectrum analyzer overlay
#   --no-vectorscope        disable stereo vectorscope overlay

Mood buckets

The compositor pre-defines six mood buckets. Clips are scored by (intensity_mean, brightness_mean) at 6-tuple ranges; matching clips get assigned to segments where both metrics fall in that bucket.

dmt_flash_editor.py — hard-cut rapid editor

Built for the Twitter-DMT-flash aesthetic: 8–12 clips, hard cuts every 1–2 s, aggressive per-beat brightness strobing, chromatic aberration, gaussian-blur bloom, vertical 720×1280 by default.

Usage

# Schedule = [{"clip_path": "...", "trim_start": 0.2, "duration": 2.0}, ...]
cat > schedule.json <<'EOF'
[
  {"clip_path": "01_neon_grid.mp4",        "trim_start": 0.2, "duration": 2.0},
  {"clip_path": "02_wireframe.mp4",        "trim_start": 0.5, "duration": 1.0},
  {"clip_path": "03_fractal_squares.mp4",  "trim_start": 0.3, "duration": 2.0}
]
EOF

python scripts/dmt_flash_editor.py \
    --schedule schedule.json \
    --audio song.flac \
    --output dmt_flash.mp4 \
    --width 720 --height 1280 --fps 24 \
    --brightness-peak 0.7 \
    --hue-deg 60 \
    --lufs -12.0

What it does

1. Loads schedule + audio
2. Detects beats via librosa
3. Builds a hard-cut concat: each clip trimmed to its specified duration, no crossfades
4. Wraps the cut video in:
   • Per-beat brightness flash (eq driven by sendcmd, +0.7 on each beat, decay over 60 ms)
   • Hue rotation by spectral centroid (hue h=... driven dynamically)
   • Chromatic aberration: rgbashift rh=+4, bh=-4 (constant, gives the "DMT pixel-shift" feel)
   • Bloom: gblur σ=8 of input + blend=screen over original (subtle glow on bright pixels)
   • Optional showcqt (spectrum analyzer overlay) + avectorscope (stereo lissajous)
5. Final loudnorm to target LUFS (default −12 for EDM)

Total typical render: ~30–60 s for a 16 s output on a modest CPU.

Tuning

# Tame the strobing
--brightness-peak 0.3   # default 0.7

# Less hue rotation (more naturalistic colors)
--hue-deg 30            # default 60

# Disable individual effects
--no-chromatic
--no-bloom
--no-cqt
--no-vectorscope

Companion workflow with aeon-music-maker

The natural pipeline is:

1. Generate the song with aeon-music-maker (see that repo's SKILL.md):

   python music_maker.py --prompt "psychedelic trance ..." --duration 90 \
       --bpm 140 --key "F minor" --variant xl_base --master edm \
       -o my_track.flac
   

2. Generate or curate clips — any video source works (LTX-rendered, stock footage, your own clips)
3. Build the music video with reactive_compositor.py (smooth) or dmt_flash_editor.py (hard-cut)

aeon-movie-maker can render the source clips for step 2 — see that repo for the LTX 2.3 22B fast cinematic pipeline.

Failure modes

See also

• aeon-music-maker — generate the audio
• aeon-movie-maker — generate the source video clips via LTX 2.3
• aeon-radio-drama — narrative audio