JEN-1
Futureverse's universal diffusion model for text-to-music, inpainting, and continuation in one network, at 48kHz stereo.
About
JEN-1 is Futureverse AI Research's answer to doing text-to-music properly: an omnidirectional diffusion model that mixes autoregressive and non-autoregressive modes end-to-end, trained on raw waveforms through a masked noise-robust autoencoder so it generates genuine 48kHz stereo without spectrogram conversion losses. One model handles text-guided generation, music inpainting, and music continuation through multi-task in-context learning.
In human evaluations the authors report 85.7/100 perceptual quality against 83.8 for the best competing methods, with strong text alignment. The paper is from August 2023 (v2 in May 2025), with audio demos on jenmusic.ai. It is a research release with no confirmed public weights or API, which keeps it in the metrics tier despite the strong numbers.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported (arXiv 2308.04729): 85.7/100 perceptual quality vs 83.8 for best baseline (human evaluation); strong text-music alignment per human judges
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (metrics tier, metrics-v1). Evidence: paper (arXiv 2308.04729, Aug 2023, v2 May 2025) from Futureverse AI Research with human-eval perceptual quality tables; demo page at jenmusic.ai/audio-demos confirmed via the paper. Strengths: unified generation/inpainting/continuation in one model, genuine 48kHz stereo waveform modeling, strong author-reported quality. Docked for: (1) evaluation is author-run against chosen baselines; (2) no confirmed public code or weight release; (3) no product, API, or independent replication. License: not confirmed in sources checked — flag for review.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 5.0/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 5.0/10
- Manual review · sanity read + red-flag check
- 6.0/10
- · Paper arXiv 2308.04729 (Aug 2023, v2 May 2025), Futureverse AI Research
- · 85.7/100 perceptual quality vs 83.8 best baseline (author-reported human eval)
- · 48kHz stereo waveform modeling; unified gen/inpainting/continuation