DiffRhythm
Full-diffusion song generation: a complete 4m45s stereo song, vocals plus accompaniment, in about 10 seconds.
About
DiffRhythm is the first end-to-end diffusion-based song generation model, turning lyrics into full-length stereo songs with both vocals and accompaniment in a single non-autoregressive pass. It pairs a sentence-level lyrics-alignment mechanism with a music VAE that shares its latent space with the Stable Audio VAE, so it can drop into existing latent-diffusion tooling. DiffRhythm-base runs in as little as 8 GB of VRAM with chunked decoding, and the code plus DiT weights are released under Apache 2.0.
Its follow-up, DiffRhythm 2 (2025), moves to a semi-autoregressive block-flow-matching design with RLHF-style cross-pair preference optimization for songs up to 210 seconds. The project stays research-led — inference is self-hosted, the maintainers flag copyright and misuse risks explicitly, and there is no official hosted API — but it is one of the few open pipelines that does true full-song generation rather than 30-second clips.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Provider-reported (DiffRhythm paper, arXiv 2503.01183): 18.2% relative PER reduction vs the SongLM baseline on song generation, with comparable musicality
- No independent third-party benchmarks published
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER metrics, metrics-v1). Evidence: Apache 2.0 code+weights (GitHub README); paper with song-generation evals (arXiv 2503.01183); active successor DiffRhythm 2 under review ICLR 2026. Strengths: fully open license, novel full-song diffusion design, honest responsible-use docs. Docked for: (1) no official hosted API or independent benchmarks — eval numbers are author-reported; (2) research-tier polish, needs 8GB+ VRAM self-host. License: Apache 2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Apache 2.0 code + DiT weights (GitHub README, verified 2026-10-16)
- · Paper arXiv 2503.01183: 18.2% relative PER reduction vs SongLM baseline
- · Successor DiffRhythm 2 (2025) under review at ICLR 2026 - active development