yue2-music-v1.2.0

minitap-ai/mobile-useyue2-music-v1.2.0Sep 25, 2026by a43992899

Upgrading?

AI Summary

A skill update for the YuE2 music generation system, adding instrumental generation and covers from text, ABC scores, or audio references.

Key Highlights

  • New instrumental generation and cover capabilities from text descriptions, ABC scores, or audio references.
  • Agent workflow integration recommended with GPT-6 Astra for running the workflow.
  • Audio references now pass through SheetSage2 for processing.

New Features

  • Support instrumental generation and instrumental covers
  • Workflow moves Vocal notes into Ins before rendering
  • Support audio references passing through SheetSage2

Full Release Notes

The **yue2-music skill now supports instrumental generation and instrumental covers** from a text description, an ABC score, or a reference recording.

**Recommended agent: GPT-6 Astra.** YuE2 writes the score and generates the audio by default; the agent runs the workflow and checks the results. Agent composition is used only when explicitly requested.

Download **yue2-music.zip**, unzip it into your agent's skill directory or import it, and ask:

> Use the yue2-music skill to create gentle piano instrumental music for reading, with no vocals. Give me playable audio and the full prompt.

For a cover, attach your recording or ABC and ask:

> Turn this melody into an acoustic-guitar instrumental cover. Keep the melody and give me the audio and full prompt.

The workflow moves every Vocal note into Ins before rendering, preserving its pitch, onset, and duration in the score. Audio references first pass through SheetSage2. The package includes setup instructions, exact model revisions, standalone checks, and local audio/prompt export. Inference requires a compatible NVIDIA GPU; weights and runtime dependencies are installed separately.

Validation: 51 instrumental CPU tests and 35 skill/packaging regression tests passed. Five fresh GPU attempts produced four complete outputs; the first ABC cover hit the generation limit and its retry completed. The strings sample had 11 full-scale samples. Listening is still needed to assess vocal leakage and audible melody fidelity. Fresh-machine installation was not tested.

SHA256SUMS covers the downloadable ZIP. No audio, model weights, credentials, machine paths, or run logs are bundled. The parent skill retains Apache 2.0; the instrumental module retains its MIT license. Model and dependency licenses remain separate.

This is a skill release. The existing YuE2 inference release is unchanged.

Related Repositories