osu!mania 4K · PyTorch · 6.97M params

ManiaMapper

AI-generated osu!mania 4K beatmaps from any song.

127
training maps
4
difficulties
90.1%
pattern acc.*

*training-set accuracy

02 · Lite demo

Map your own song

Pick a track and a difficulty. Everything runs in your browser — your audio never leaves your device.
Simplified rule-based heuristic — not the neural network

Difficulty

Detected BPM

—

Total notes

—

Notes / sec

—

Long notes

—

What the Lite demo does

Spectral flux on 1024-sample frames finds onsets, autocorrelation estimates the BPM, and notes are placed on a 1/4-beat grid above an adaptive threshold tuned to the difficulty's notes-per-second target. Chords go on the strongest hits, long notes on sustained loud passages, and columns are balanced with a 60 ms same-column gap.

Unranked: your own song

Load a song to preview its map

0:000:00
Play mode

This is a simplified rule-based version. The real ManiaMapper uses a neural network trained on 127 human-made maps to learn patterns, not just onsets.

03 · The real model

How ManiaMapper works

A PyTorch BiLSTM audio encoder and a causal LSTM pattern decoder — about 6.97M parameters — turn a song into a playable map.
  1. 01

    Audio file

    any song
  2. 02

    Librosa analysis

    BPMbeatsonsetsmel spectrogramchromaspectral features
  3. 03

    Beat grid

    time steps aligned to beats
  4. 04

    Encoder + Decoder

    mania_model.ptdifficulty-conditionedautoregressive
  5. 05

    Note assignment

    column selectionmin-gaprebalancing
  6. 06

    Speed variation

    optional
  7. 07

    .osu writer

    timing + hit objects
  8. 08

    .osz package

    ready to import

Architecture

~6.97M parameters
Input

(B, T, 130) = 114 audio features + 16 cross-difficulty context

Difficulty embedding: 4 levels → 32-dim

Encoder

2 × multi-scale depthwise conv blocks, kernels 3 / 5 / 9 / 13 / 25

4-layer BiLSTM, 384-dim

Decoder

Causal 2-layer LSTM, 256 hidden

+ 64-dim embedding of the previously placed event

Heads

Per difficulty: hit (4 columns) · LN start (4) · LN duration (4)

+ 18-class pattern head

04 · Training results

What it learned from

One documented training run over 50 epochs on human-made 4K maps.

Usable maps

127 / 151 scanned

Easy 14 · Normal 21 · Hard 52 · Insane 40

Training sequences

2,408

of 256 steps each

Regular notes

222,013

hit objects

Long notes

45,807

hold objects

Total loss per epoch

15.64 → 0.30 · log scale

Pattern accuracy per epoch

7.9% → 90.1% · training set

Difficulty presets

LevelHPODMax chordTarget npsTarget LN %
Easy6612.54%
Normal7724.57%
Hard8829.510%
Insane99316.012%

Honest limitations

  • 90.1% is training-set accuracy — there is no held-out validation split.
  • There's no automated playability metric, so map quality is a human judgement.
  • 4K only.
  • A single documented training run.