Repository navigation
Keyflow 1.6.0 · Human-chart-trained levels
Keyflow now prepares complete learned charts in the background before a song becomes playable. Opening a prepared level reads its saved tile locations and highlight without running chart inference.
- Replaces the handwritten tile-pattern planner with the MIT-licensed Mapperatorinator-v32-mini mania checkpoint, trained upstream on community osu! charts and adapted for local Android inference.
- Predicts note times, lanes and hold boundaries from audio and earlier chart context. Keeps the two-thumb safety gate and varied replay layouts.
- Runs both rhythm and chart models on the phone; no server, API key or audio upload. Preparation progress is shown while music becomes ready.
- Covers audible song intros that the pretrained model would otherwise leave empty.
- Saves full charts and journey highlights atomically; interrupted preparations can retry, and existing audio is reused for chart version 6.
Validation: JVM tests, Android lint, model checksums, release-tool tests and native Android checks. See the included chart-learning documentation for measured results, reproducible conversion and limitations. This release adapts published pretrained weights; it does not claim new custom training or hand-authored quality for every song.
The larger APK includes both neural models and the existing offline runtime. Development releases may reset app data. APK, source archive and SHA-256 checksums are attached by the release workflow.