SongBloom
Open lyrics-to-song model: full songs with vocals and accompaniment from lyrics plus a 10-second style prompt.
About
SongBloom is a lyrics-to-song generation model that produces full-length songs, vocals and accompaniment together, from structured lyrics and a 10-second reference audio clip. It uses an interleaved design: an autoregressive pass sketches the song's structure and a diffusion stage refines the acoustics, which keeps long songs coherent instead of drifting. Two sizes were released, a 1.3B tiny model for 60-second songs and a 2B full model for songs up to 150 seconds at 48kHz, with a DPO post-trained version of the full model arriving in September 2025.
The paper (arXiv 2506.07634, accepted at NeurIPS 2025) reports wins over SongGen, SongEditor, DiffRhythm and YuE on both objective and subjective measures, and claims quality comparable to commercial song-generation platforms. Those comparisons are provider-reported, so treat them as a signal rather than proof. Code and weights are out under Apache 2.0 on GitHub and Hugging Face (CypressYang/SongBloom), which makes it one of the few genuinely open full-song generators.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported wins over SongGen, SongEditor, DiffRhythm and YuE on objective and subjective metrics; no independent third-party benchmark found.
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: peer-reviewed paper at NeurIPS 2025, open Apache-2.0 code and weights, active repo with a DPO update in Sep 2025. Strengths: rare open full-song generator, coherent long-form structure, honest 48kHz output. Docked for: (1) all quality comparisons are provider-reported with no independent replication; (2) young project with a short maintenance history. License: Apache 2.0 (per project README).
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
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · NeurIPS 2025 paper (arXiv 2506.07634); 2B full model, 150s songs at 48kHz
- · DPO post-trained full checkpoint released Sep 2025
- · Provider-reported wins over SongGen, SongEditor, DiffRhythm, YuE; no independent benchmark
- · Open code and weights on GitHub/Hugging Face