TuneJury — instance-level pairwise reward model for text-to-music

TuneJury is a 2.8M-parameter MLP head over frozen LAION-CLAP-Music + MERT-v1-330M embeddings, trained with a shared-weight pairwise-logistic objective on ~17.5K human A vs. B preferences from four open sources. It scores a single audio clip (with an optional text prompt) as one preference scalar.

• Paper: https://arxiv.org/abs/2606.17006
• Project Page: https://yonghyunk1m.github.io/TuneJury/
• Code: https://github.com/yonghyunk1m/TuneJury
• Demo: https://huggingface.co/spaces/TuneJury/tune-jury-demo
• Scope: trained and evaluated on instrumental music. For prompt formats unlike the live-arena style it was trained on (e.g. post-hoc captions), an empty prompt can be the safer default (paper §4.2). Anchor calibration adapts to newly released generators without retraining.

Checkpoints

Quick scoring

Install the package (also pulls the LAION-CLAP encoder ~2.2 GB on first use; needs ffmpeg and libsndfile):

pip install git+https://github.com/yonghyunk1m/TuneJury

from huggingface_hub import hf_hub_download
from tunejury.score import Scorer

sc = Scorer.from_pretrained(hf_hub_download("TuneJury/tunejury", "tunejury.pt"))
print(sc.score("clip.wav", ""))   # "" -> 512-d zero text vector (paper §3/§4.2 empty-prompt)
print(sc.score("clip.wav", "a calm lo-fi piano loop"))   # a prompt uses the text branch

License

The released head is CC-BY-NC 4.0, tracking the strictest upstream weight license (MERT-v1-330M). Frozen encoders at inference: LAION-CLAP-Music (CC0 1.0), MERT-v1-330M (CC-BY-NC 4.0), MuQ-MuLan-large (CC-BY-NC 4.0). The A1_clap_audio_only head is released under Apache-2.0.

Citation

@misc{tunejury2026,
  title         = {TuneJury: An Open Metric for Improving Music Generation Preference Alignment},
  author        = {Kim, Yonghyun and Lee, Junwon and Xia, Haiwen and
                   Ma, Yinghao and Koo, Junghyun and Saito, Koichi and
                   Mitsufuji, Yuki and Donahue, Chris},
  year          = {2026},
  eprint        = {2606.17006},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
  url           = {https://arxiv.org/abs/2606.17006},
}