DiffRhythm
Open-source diffusion model that writes full songs with vocals in seconds
About
DiffRhythm is the first open-source diffusion-based model for full-length song generation, from researchers at Northwestern Polytechnical University. The DiffRhythm-full release (March 15, 2025) generates complete songs up to 285 seconds — vocals and accompaniment together, end to end — from lyrics plus a style prompt, in as little as ten seconds. It is Apache 2.0 licensed (since March 7, 2025), supports English and Chinese lyrics, text-based style prompts, instrumental mode, and song editing/continuation (v1.2, May 2025). Being diffusion-based rather than autoregressive, it offers better controllability and editing (inpainting, style transfer) than LM-based song models. Practical notes: output quality depends on your GPU; arrangement richness improved markedly between v1 and v1.2; and like all generative music, commercial use needs care around training-data provenance.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported: full song in ~10 seconds end-to-end
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Open-source under Apache 2.0 since 2025-03-07; paper arXiv:2503.01183; DiffRhythm-full release 2025-03-15: 285s songs with vocals in ~10s, team-reported; v1.2 (2025-05) improved arrangement and editing; reproducible weights.