You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access to the YTSV token releases is granted for non-commercial research use. These tokens are derived from YouTube score-following videos whose underlying media remains the property of its respective uploaders; only discrete token representations are distributed here. The YTSV metadata this derives from is licensed CC BY-NC-SA 4.0, so the non-commercial and share-alike terms travel to these tokens and to anything derived from them. Requests are reviewed manually; please describe your intended use so it can be assessed.
Log in or Sign Up to review the conditions and access this dataset content.
YTSV score-image tokens (U-MusT, unirqvae3)
RQ-VAE score-image tokens for the YouTube Score Video (YTSV) dataset, the main dataset contribution of U-MusT — 433,920 image–audio pairs from 12,217 score-following videos, 1,341 hours in total, covering roughly 10,000 pieces by more than 2,000 composers. An order of magnitude larger than any prior music modal-translation corpus.
Tokenized with unirqvae3-ytsv at 16x compression. Roughly 100 GB across 1.1M token files before sharding.
No video or audio is redistributed. These are tokens only. The video metadata and the preprocessing pipeline that produced them live in MALerLab/youtube-score-video-dataset; the media itself must be downloaded from YouTube by you.
Layout
The token tree is sharded into one gzipped tar per collection group, because the uncompressed form runs to over a million small files — impractical for a Hub repository to serve.
image/0-2.tar.gz
image/0-3.tar.gz
... 18 shards, ~5.6 GB each
Each archive expands to <group>/segments/<youtube_id>/image_tokens/..., so extracting all
shards under one directory reconstructs the layout the U-MusT loaders expect:
hf download malerlab/ytsv-unirqvae3-ytsv --repo-type dataset --local-dir shards/
mkdir -p dataset/latent_score_dataset_tokens
for f in shards/image/*.tar.gz; do
tar -xzf "$f" -C dataset/latent_score_dataset_tokens
done
Split manifests for the training, validation and test partitions are in the
code repository under dataset_pair_paths/
(lsyt.json.gz for all instrumentation, lsyt_piano.json.gz for the piano subset).
Token files contain shift variants, not a single tokenization
Read this before loading anything. A one-pixel offset in a score image changes the discrete token assignment almost entirely, so training on a single tokenization makes a model brittle to offsets it will certainly meet at inference. Every file here instead stores a stack of discretization-shift variants on its leading axes.
YTSV image tokens are .pt files of torch.int16 with shape (8, 4, H, W, 4):
| Axis | Size | Meaning |
|---|---|---|
| 0 | 8 | horizontal one-pixel shifts |
| 1 | 4 | vertical one-pixel shifts |
| 2–3 | H, W |
token grid of one musical system at 16x compression |
| 4 | 4 | the RQ-VAE's four codebooks |
Index [0, 0] is the unshifted tokenization. These come from real score videos rather than
synthetic renderings, so there is no key-transposition axis — the (6, 8, 4, H, W, 4) shape
found in synthetically rendered corpora does not occur here.
Full details of the augmentation scheme are in Appendix B of the supplementary material, published with the article at doi.org/10.1109/TASLPRO.2025.3648794.
Codec compatibility — read before pairing with a checkpoint
Image tokens are not interchangeable between codec generations. The unirqvae and
unirqvae3 codebooks differ, so the same token index means different things. Pairing a
checkpoint with tokens from the other generation produces meaningless output, silently.
| Released run | Direction | Image codec it expects |
|---|---|---|
run-20250225_062905-9n1554as |
Image-to-Audio, piano | unirqvae3 |
run-20250130_150202-x9znhap2 |
Image-to-Audio, multi-instrument | unirqvae |
run-20250128_025927-ks0ibl4v |
Audio-to-Image, multi-instrument | unirqvae |
Every published image-token dataset in this collection is unirqvae3, so only the piano
Image-to-Audio run pairs with them directly. For the strings and Audio-to-Image runs, image
tokens must be re-baked with unirqvae-ytsv
using scripts/bake_image_tokens.py from the
code repository. This matters most for the
Audio-to-Image direction, where image tokens are the output target — evaluating it against
unirqvae3 tokens measures nothing.
infer.py reads the tokenizer from each run's own files/config.yaml, so inference on a
downloaded checkpoint selects the right codec automatically. The mismatch only arises when
pairing a checkpoint with a token dataset by hand.
Licensing
Released under CC BY-NC-SA 4.0, matching the license of the YTSV metadata it derives from.
The underlying videos are the property of their respective uploaders and are not covered by this license. This repository distributes only discrete token representations computed from them. The full per-corpus license table for the project, and the conflicts between corpora, are documented in the code repository. This is not legal advice — satisfy yourself that your intended use is permitted.
Citation
@article{jung2026umust,
title = {U-MusT: A Unified Framework for Cross-Modal Translation of Score Images, Symbolic Music, and Performance Audio},
author = {Jung, Jongmin and Kim, Dongmin and Lee, Sihun and Cho, Seola and Soh, Hyungjoon and Bukey, Irmak and Donahue, Chris and Jeong, Dasaem},
journal = {IEEE Transactions on Audio, Speech and Language Processing},
volume = {34},
pages = {1876--1891},
year = {2026},
doi = {10.1109/TASLPRO.2025.3648794}
}
- Downloads last month
- 17