malerlab/u-must-contin-u

Contin-U: a whole PDF piano score → one continuous performance, via the two-system sliding window of the U-MusT image-to-audio model (RenCon 2025 / MIREX).

Public
7 runs

Run time and cost

This model runs on Nvidia L40S GPU hardware. We don't yet have enough runs of this model to provide performance information.

Readme

Contin-U · Full PDF piano score to one continuous performance

Contin-U turns the U-MusT image-to-audio model, trained on windows of 1–3 systems, into a full-score renderer without retraining. The PDF is rasterized, every musical system is detected in reading order, and the model slides over consecutive system pairs (j, j+1): the decoder is primed with the audio already generated for system j, the cross-attention of a late decoder layer onto the tokens after the [SEP] separator marks where system j’s audio ends, the stream is spliced there, and the tokens that follow prime the next window. The result is a single continuous performance, decoded once from the joined audio tokens.

Contin-U was the MALerLab entry to RenCon 2025, the expressive piano performance rendering contest revived as a MIREX task at ISMIR 2025 (Daejeon). Paper: Contin-U: Full-Score to Performance Audio with Cross-Attentive System-Continuation Inference (Jung, Kim, Lee, Cho, Soh, Bukey, Donahue, Jeong) — https://futuremirex.com/portal/wp-content/uploads/2025/rencon/Contin-U.pdf

Inputs

name description
score PDF of a piano score (engraved or scanned).
first_page, last_page Page range (1-based; last_page = 0 = to the end).
dpi Rasterization resolution, 150–300.
max_systems Stop after this many systems; 0 = all.
attention_threshold Cross-attention mass on the second system that marks the boundary (default 0.5).
seed Random seed.

Outputs

A list of files (order may vary): contin-u.wav (the whole performance), meta.json (duration_sec, n_pages, n_systems, notes), systems.png (all system crops stacked in playback order) and pPP_sSS.png (each crop, labelled with its position, page and system).

Notes

  • One window (two systems) takes about 30 s of GPU time on an L40S; a 6-system page is 5 windows.
  • Piano only; the model expects grand-staff systems. Title pages without music are skipped automatically.
  • Research use under CC BY-NC-SA 4.0.

Code: https://github.com/MALerLab/U-MusT · U-MusT paper: https://doi.org/10.1109/TASLPRO.2025.3648794

Model created
Model updated