M.I.C.D.R.O.P. Concept
M.I.C.D.R.O.P. is a research initiative to build a neural composer that generates FNF-style instrumentals and vocals from text prompts BPM-accurate, mod-ready, and purpose-built for rhythm game music.
M.I.C.D.R.O.P.
Machine-Intelligent Composer for Dynamic Rhythm and Original Performance.
The concept: type a prompt, get back a rhythm-game-ready track. Not a generic music model a system specifically trained on the structure, energy, and format of FNF music.
What it generates
Given a prompt like "Tiffany vs Cassian 175 BPM ragecore duet", the system outputs:
instrumental.wavvocals.wavchart.json(BPM, section structure, optional note mapping)
Outputs are synchronized stems ready for mod engines or DAWs.
Technical approach
Text-conditioned transformer architecture (similar to MusicGen), trained on tokenized audio via EnCodec. Text conditioning through CLAP or T5. Stem separation via Demucs post-generation.
Training data is sourced from FNF mod assets paired song.ogg / voices.ogg stems with chart metadata providing BPM, section type, and character context.
Current phase
Dataset construction scripted extraction of mod assets, audio normalization, and caption generation. Training targets a high-end GPU (5080-class).
This is early-stage research. The concept is clear; the dataset work comes first.