About this music
Every track on this station is composed, performed, illustrated and mastered by artificial-intelligence models running on a single local machine — no studio, no session musicians, no stock library. A style prompt and lyrics go in; a finished song with cover art and a looping Canvas video comes out.
Models & tools
- Song & score — YuE2-3B (m-a-p)
- Open lyrics-to-song model: it writes the vocals, the arrangement and the editable ABC score. Served locally through the community GGUF build with the audio.cpp runtime. Model weights are CC BY-NC 4.0 (free for personal and creator use; commercial use needs a licence from the authors).
- Instrumentals — community AR LoRA
- Vocal-free tracks (game assets, background music) use an instrumental adapter for YuE2 trained by the community. The base model always sings; the adapter teaches it not to.
- Cover art — Qwen Image
- Square album artwork is diffusion-generated from each song's style and lyrics.
- Canvas video — MiniMax video model
- The vertical 9:16 loops are image-to-video clips anchored on the cover art, then seamlessly looped for endless playback.
- Songwriting ideas — local writer LLM
- Batch themes, titles and lyric drafts are rolled by a local language model inside pinned style packs, so each song varies within its genre.
- Mastering — FFmpeg, EBU R128
- Every track is loudness-normalised to −14 LUFS broadcast standard and encoded to AAC and Opus for universal playback, with a precomputed waveform for the scrubber.
Hosting
Audio, artwork and video are served from Cloudflare R2 object storage (zero-egress) and this player is a static site on Cloudflare Pages. Made with Local AI Studio — a fully local voice and music studio.