MusicGen

Open weights you can run locally, short clips steered by a melody.

Model profile

Vendor

Meta

Form

本地模型

Input

文本 · 旋律 · 参考音频

Output

音频

Licence

代码与权重分别核对，权重的商用范围以模型卡与随附许可为准

Verified

2026-10-30

Related

MelodyMIDI ProtocolPiano

What it lets you finish

Generate a few dozen seconds of instrumental audio from a text description, or hand it a melody and ask for a given timbre and style around that line. The weights can be downloaded and run offline, which suits batch auditioning, building a sample library, and further training on your own audio. It returns clips, not songs; whole-form structure is your job or another tool's.

Inputs and outputs

Input is text, optionally with an audio clip as a melodic condition so the output follows that line. Output is an audio file with no stems and no MIDI. Running locally, you control sampling parameters, duration and seed, and one seed reproduces one result. The listening examples and instances below are references for this kind of output, not this model's generations.

Strengths

• Local execution keeps material on your machine, which suits confidential work.
• Melody conditioning works: give a hummed or single-note line and the arrangement follows it, which speeds up arrangement sketches.
• You can continue training on your own audio to approach the timbre of your own catalogue.
• Batch generation is cheap, so running dozens of candidates and selecting is practical.

Weaknesses and failure modes

• It does not sing. Vocal requests become indistinct vocal-like texture, so it is not a song tool.
• Clips are short and start circling past half a minute: one motif repeats and the ending does not close. A reproducible test is two runs from one prompt; both start repeating around a similar point.
• Over-specific melody conditioning makes the accompaniment track every note, flattening the harmony into a scale-following part.
• Text descriptions of instrumentation are half ignored; ask for strings and electric piano and often only one appears.
• Output level and stereo image vary, so a batch needs per-file work before it enters a session.
• Default quality trails hosted services: muffled highs and soft cymbals are common and need post work.

Terms, price and alternatives

Confirm: code licence and weight licence are separate questions, and whether weights may be used commercially or to train derivatives belongs to the model card and its accompanying licence; training data provenance follows the official statement. Local runs carry no per-use fee, but you pay in compute. Alternatives: Stable Audio for out-of-the-box quality and longer clips; DiffSinger or ACE Studio when you need singing. This is the option that you can modify, retrain and run offline.

Sources

• 模型能力、条件输入与权重发布形态，依官方模型说明核对，未抄原文
• 失败模式依公开用户报告与本地实测经验归纳，未引用具体评测数据

Used for fact-checking only; all prose is original.