0MGE

0MGE: Neural Granular Engine

AI music generation from YOUR music. Scan, train, generate — all locally. No cloud, no API, no subscription.

Pre-trained neural granular engine trained on 2389 tracks (~48 hours of music). Generates new drone landscapes, textures, and atmospheres from a learned grain field.

Full source code, training pipeline, and desktop app: 0MGE on GitHub

Listen to demo — audio-reactive player with GLSL shader

---

What This Is

A neural network that learns from music and generates new sound. Not text-to-music. It scans audio files, cuts them into millions of micro-grains, and trains a navigator to assemble those grains into new sonic worlds.

This repo contains a pre-trained model and grain pool trained on Slut Online's music (2389 tracks). Download, generate, hear what it sounds like.

To generate from your own music — use the desktop app (scans your library, trains locally).

---

Generate

git clone https://github.com/0penAGI/0MGE.git && cd 0MGE
pip install numpy torch librosa scikit-learn soundfile

python3 granular_field.py --pool granular_pool_v2_int16.npz --model granular_multi_v1.pt --bars 60 --multi-stream --seed 42

Output: granular_output/granular_60bars_*.wav (stereo, 22050 Hz).

---

Files

---

Architecture

MultiNavigator — Transformer (4 heads, 3 layers, 192 hidden). 48-dim state, 12-step context. 6 independent stream heads select grains via softmax over pool similarities. Attractor field (z0-inspired) provides learned global state per stream for long-range coherence — pulls generation toward meaningful musical directions instead of random walk.

---

Grain Pool

Three-tier hierarchy extracted via STFT (n_fft=1024, hop=256):

22-dimensional spectral features per grain. MiniBatchKMeans clustering (1024 clusters).

---

Quantization

---

Audio Samples

---

Citation

@software{0mge2026,
  title={0MGE: Neural Granular Engine},
  author={0penAGI},
  year={2026},
  url={https://github.com/0penAGI/0MGE}
}

License

MIT

Trained on music by Slut Online with permission.