Back to Gap Map Model components / Image, video, 3D & music generation

DiffRhythm

ASLP Lab
restricted / Overall score: 1.0

DiffRhythm is a song generation model line from ASLP Lab at Northwestern Polytechnical University. It writes a complete song, vocals and accompaniment together, from lyrics and a style prompt or reference clip. DiffRhythm 2, released with Xiaomi in October 2025, generates semi-autoregressively in blocks to keep the singing aligned with the lyrics.

The first DiffRhythm (March 2025) and its 1.2 update remain on the Hub with training code. DiffRhythm 2 was developed with Xiaomi and ships inference code only; it loads the separately published MuQ-MuLan style encoder.

Openness

2 medium confidence
2.0
weights
open (the DiffRhythm 2 checkpoint and decoder download from the Hub without a gate)
data
closed (no training songs are released)
code
partial (DiffRhythm 2 publishes inference code only; training code was published for the first DiffRhythm)
license
Apache-2.0 (the DiffRhythm 2 code and weights)+CC-BY-NC-4.0 (the MuQ-MuLan-large style encoder the inference script downloads and loads on every run)

DiffRhythm 2's own code and weights are Apache 2.0, but generating a song requires the MuQ-MuLan style encoder, which its authors license for non-commercial use only, so the model as run cannot be used commercially. Only inference code is published for this release, and the training data is not released.

Adoption

1 high confidence
1.0

Adoption is measured as Hugging Face downloads summed across ASLP Lab's DiffRhythm checkpoints, most of them DiffRhythm 2.

Capability

1 medium confidence
1.0

DiffRhythm 2 generates full songs with vocals but has no placement on the public music arenas, where MusicGen at least appears.

Verified 2026-11-12