Licensing and provenance

Why you need this

Running a model on your own machine does not make its output yours to use however you like. The model came with a license, and the law on owning machine-made work is still being worked out. Before a file ships, you need to know what the model's license allows, and you need a record that says exactly how the file was made. This page is the last step of stage 5 of the pipeline. It explains what the documents say; it is not legal advice, and a game with real money behind it should have a lawyer read the licenses.

The idea

Two separate questions.

1. The model license: what the people who released the model allow you to do with their code and weights (the stored numbers that make up the trained model). This is a contract-like grant, and it can put conditions on how you use the model, including what the use is for.
2. Rights in the output: whether anyone owns copyright in the clip that came out, and who. That is a question of national law, not of the license, though a license can add promises or conditions about outputs.

Open weights versus open source. "Open weights" means the weights can be downloaded. It says nothing about what you may do with them. "Open source", in the sense the Open Source Initiative defines, requires that the license not restrict fields of use, so a license that forbids commercial use is not open source under that definition, even if it is short and looks like a familiar one. Read the conditions, not the label.

What the two music models' documents say, as published when this page was written:

Model	License document	Clause that matters	What it means for a shipped game
ACE-Step 1.5	The MIT License, in the LICENSE file of the ACE-Step 1.5 repository; the Hugging Face model card's metadata also declares MIT	The model card says: "You can strictly use the generated music for commercial purposes."	The license permits use, modification, and sale of the software; the model card states that outputs may be used commercially.
LeVo 2 (SongGeneration 2)	The "License Terms of SongGeneration" in the LICENSE file of the SongGeneration code repository, which covers "the SongGeneration inference-enabling code and the weights"	"You agree to use the SongGeneration only for academic, research and education purposes, and refrain from using it for any commercial or production purposes under any circumstances."	A released game, paid or free, is hard to describe as academic, research, or education use. Read this clause before you plan on LeVo 2 for anything you ship.

Four notes on the table. The SongGeneration license says nothing about who owns outputs; its condition is on how you use the model. The Hugging Face page for the LeVo 2 large weights declared no license in its metadata when checked, so the code repository's LICENSE file is the document to read. Tencent's official repository and model card could not be read when this page was written, so the clause above is quoted from a copy of that repository made after the LeVo 2 release. Check the license on the model's own page at the time you download it, and rely on what you read there. And the same repository's README describes LeVo 2 as achieving "commercial-grade generation", which describes its quality, not permission: the README and the license sound different, and the license is the one that grants rights.

The ACE-Step 1.5 repository also carries a disclaimer. It names "unintentional copyright infringement due to stylistic similarity" as a risk and encourages users "to verify the originality of generated works, clearly disclose AI involvement, and obtain appropriate permissions when adapting protected styles or materials." That is a request, not a license condition, but it is good practice either way.

Copyright in generated work is unsettled. It varies by country and is changing. One primary source, as an example of where one country stands: the United States Copyright Office published part 2 of its Copyright and Artificial Intelligence report, on copyrightability, in January 2025. The Office's announcement of it states: "The outputs of generative AI can be protected by copyright only where a human author has determined sufficient expressive elements." It adds that this "can include situations where a human-authored work is perceptible in an AI output, or a human makes creative arrangements or modifications of the output, but not the mere provision of prompts." Other countries take other positions, and courts are still deciding cases. The practical upshot for a game maker: you may not be able to stop someone else from reusing a raw generated clip, and your own edits, arrangement, and lyrics are the parts most likely to count as yours.

Do not imitate a named artist. A prompt like "in the style of" a living artist, or lyrics borrowed from a real song, raises the chance that the output resembles protected work. Describe genre, mood, instruments, and tempo instead.

Provenance is the record of where a file came from: which model and version, which seed and settings, when, by whom, and every edit after. With it, you can answer a question about a file a year later, regenerate it, or replace it if a license changes.

Worked example

The values are illustrations, not a record of a real run. This is the provenance record for the Lumen Clash lobby loop, built on the sibling pages: generated with ACE-Step 1.5, looped on Seamless music loops, leveled on Loudness for games, exported on Exporting game audio.

{
  "file": "music/lobby-loop.ogg",
  "master": "masters/lobby-loop.flac",
  "master_sha256": "(computed from the master file)",
  "model": "ACE-Step 1.5",
  "model_variant": "acestep-v15-sft",
  "model_license": "MIT (repository LICENSE); model card states generated music may be used commercially",
  "license_checked_on": "2026-10-02",
  "prompt": "cinematic fantasy, warm felt piano, soft synth pads, round electric bass, light percussion, polished modern production, calm, hopeful, spacious, gentle",
  "lyrics": "[Instrumental]",
  "take": "t03",
  "settings": { "seed": 4129, "duration": 48, "bpm": 120, "keyscale": "A minor", "inference_steps": 50, "guidance_scale": 7.0 },
  "settings_file": "workflows/lobby-loop-t03.json (every other setting, saved with the take)",
  "generated_on": "2026-10-01",
  "generated_by": "the person who ran it",
  "edits": [
    "loop start 192000, loop end 960000 samples at 48 kHz (bars 3 to 10)",
    "960-sample equal-power crossfade into the audio before the loop start",
    "gain -2.0 dB, from -16.0 LUFS to -18 LUFS; true peak -3.5 dBTP"
  ],
  "export": "Ogg Vorbis, 48 kHz stereo, nominal 128 kbps; AAC copy for Safari",
  "artist_references_in_prompt": "none"
}

Each field answers a question someone will ask:

Question	Field
Can we still use this if a license changes?	model, model_license, license_checked_on
Can we make it again?	prompt, settings, settings_file, model, model_variant
Is the shipped file the one we think it is?	master_sha256
What did a person add?	edits
Did we avoid imitating anyone?	artist_references_in_prompt

Keep one record per shipped file, next to the masters, under version control. A single folder of these files is your answer to "where did this sound come from?"

In a game's audio pipeline

This is the last stop before release. The record starts in stage 3, when curating takes logs prompt, seed, settings, and model version for every keeper, and it grows through stage 4 as each edit is made. Choosing a local audio model is where the license should first be read: choose the model for a shipped asset knowing what its license says, not after the music is finished.

Common mistakes

• Treating "downloadable" as "allowed". Symptom: a finished soundtrack from a model whose license excludes commercial or production use.
• Reading the license once. Symptom: the record says MIT, but the current license file says something else. Record the date you checked.
• No seed or settings kept. Symptom: a bar needs fixing and the take cannot be reproduced.
• Edits not recorded. Symptom: the master and the shipped file differ and nobody knows how.
• Artist names in prompts. Symptom: a clip that sounds a little too much like a specific real song.
• Assuming you own the raw output. Symptom: surprise when a generated clip turns up in someone else's project and there is nothing to be done.

Cost

No generation time and no shipped bytes: a provenance record is a few hundred bytes of text per file. The cost is the maker's time, a minute or two per file when the take log already has the data, and much longer when it has to be reconstructed after the fact. Reading two model licenses carefully takes less time than one regenerated soundtrack, which is what a missed clause costs.

Going further

• The full license text and model card of every model you use, reread at each release.
• The United States Copyright Office's Copyright and Artificial Intelligence report, parts 2 (copyrightability) and 3 (training).
• Your own country's copyright office guidance on machine-generated works.
• The Open Source Initiative's Open Source Definition, for what "open source" requires of a license.