How it works

A small model, a foot model, and your CPU.

StepForge Flow is the charting model inside StepForge. It is small enough to run on any recent computer in seconds and was built to write charts people can actually dance. Here is what it does, how well it does it, and what it learned from.

Signal processing

1. Listen

Finds the song's tempo, beat grid and every onset, and how its energy moves from section to section. This is classic audio analysis, not a neural network.

RhythmNet · 9.6M parameters

2. Place

Reads the music on a musical grid of 48 ticks per bar and decides where steps, jumps, freezes and mines go: real 4th, 8th, 16th and triplet positions, busier in choruses and calmer in verses.

ChoreoNet · 15.2M parameters

3. Step

Chooses the arrows. It thinks in terms of feet ("the other foot steps on its home panel") rather than arrow columns, which is why its patterns flow.

Foot model

4. Check

An exact search over every possible foot position picks the final arrows: no impossible spins, no stepping with a foot that is holding a freeze, no unforced double-steps, crossovers and footswitches only where the difficulty allows them.

Both networks run with ONNX Runtime on the CPU; there's no PyTorch and no graphics card involved. The full write-up is the model card.

Results

How good are the charts?

180 songs with human-made charts, none of them seen in training. StepForge charted each one blind from the audio, and we compared it step by step with the human chart of the same song and difficulty.

180 songs it never sawStepForge Flow v1.4Best open-source charter we testedTwo humans
Steps placed where the human charter put one (F1, ±20 ms)0.6560.674≈ 0.75
… Beginner0.5190.587
… Hard / Challenge0.714 / 0.7260.703 / 0.712
… on the 120 songs never used for tuning0.6450.671
Chart statistics inside the human range98%90%94%
Unforced double-steps0.01%1.04%0.35%

Step placement is level with the best open-source tool overall; the Hard and Challenge lead is not statistically significant, and Beginner is clearly behind. Playability is where StepForge stands out. Highlighted cells are clear wins.

Reading the table

F1 scores how many steps land where the human charter put one (within 20 ms): 1.0 would be identical. Two different humans charting the same song agree at about 0.75, so that's roughly the ceiling. Differences were tested with a paired bootstrap over songs; only the cells highlighted green are clear wins.

No tuning on the test

Some of StepForge's decoding settings were tuned on 60 of the 180 benchmark songs. On the other 120, which nothing was tuned on, the picture is the same: 0.645 against 0.671 for the best open-source tool, a difference that isn't significant.

Memorization check

Does it copy its training charts? No.

A model trained on charts could, in principle, learn to reproduce them. We tested whether StepForge does, two ways.

0.733 vs 0.734

Its own training songs

Charting the 227 songs it trained on, it matches the human charts no more closely than on 27 songs from the same packs that it never trained on. Its longest exact copy of a training chart is 9 steps. For contrast, two human charts of the same song share runs of up to 366 steps.

0.016% vs 0.26%

New songs

Of StepForge's 16-step sequences on the 180 new songs, 0.016% appear anywhere in its training charts, against 0.26% for the human charts of those songs. No 32-step sequence appears at all, from either.

Method and caveats: model card, memorization check.

Training data

254 community songs, every one credited.

StepForge Flow v1.4 learned from 254 songs and 1,066 charts (about 8 hours of audio).

StepMix 1.0, 2, 3, 4 and 5 (community contest packs)235 songs
OutFox Serenity, Volumes 1–317 songs
StepMania 5 default songs2 songs

Every song is listed in CREDITS.csv with its artist, step artists, pack, audio and chart license, the attribution text and where it came from. Thank you to all of them.

Licenses vary by song. Most are StepMix contest entries under the contest's grant; the rest carry Creative Commons licenses or the OutFox Serenity free-content terms. The songs were checked for redistribution permission, but most of these licenses don't address machine-learning training one way or the other. We'd rather be upfront about that than pretend otherwise. It's why the opt-out below exists, and why the model's weights carry the most restrictive terms of their training data (non-commercial, share-alike).

None of StepForge's own charts were used to train it.

Known limits

Where it falls short today.

  • Beginner and Easy are the weak tiers. On about 1 song in 8 the beat grid lands half a beat off, and those songs' easiest charts end up between the beats. Beginner and Easy charts also have more notes than human ones (about 1.65× and 1.4×).
  • About a fifth of a human chart's notes are choices another charter would have made differently, so expect taste to differ.
  • 4-panel (dance-single) charts only, in 4/4, at one constant tempo per song. No doubles, no gimmicks, and songs with big tempo changes won't sync well yet.
  • This is one model from one training run, on a fairly small dataset. Opt-ins are how future versions can learn from more.

Licenses & disclosure

Source code: MIT. Model weights: CC BY-NC-SA 4.0 (non-commercial). Charts you make are yours. Every chart StepForge writes is credited “StepForge (AI-generated)”; if you share them, please keep that credit.