Skip to content

Music ​

The model you name decides how a song is made; the three kinds differ in output and in whether lyrics are required.

Available music models ​

ModelPipelineOutputSizePartsRegistries

The same catalog dppan models --music reads; it also shows what is installed and, for a partial install, what is missing.

OutputLyricsSeparate tracks
minimax44.1 kHz stereooptionalno
acestep48 kHz stereooptionalno
yue16 kHz monorequiredvocal / instrumental

Read the sample rate and channels from the response rather than assuming them.

Getting a music model ​

A music model is four or five files in one directory. MiniMax-Music3:

bash
BUNDLE=~/.dppan/models/bundles/minimax-music3

# The language model comes from a different repository.
dppan pull hf:Serveurperso/MiniMax-Music3-GGUF:Q8_0 --component language_model    --out $BUNDLE

# The other four.  --f16 halves the transformer; see below.
dppan pull hf:audio-cpp/MiniMax-Music3-GGUF:BF16 --component transformer --f16    --out $BUNDLE
dppan pull hf:audio-cpp/MiniMax-Music3-GGUF:Q8_0 --component rvq_depth_decoder    --out $BUNDLE
dppan pull hf:audio-cpp/MiniMax-Music3-GGUF      --component condition_encoder    --out $BUNDLE
dppan pull hf:audio-cpp/MiniMax-Music3-GGUF      --component vocoder              --out $BUNDLE

dppan-orchestrator

About 20 GB, or 15 GB with --f16. The orchestrator serves it on its next start.

  • Pass the same --out every time. Without it, each pull lands in a directory named after its own repository, and no directory holds a whole model.

  • Component names are the publisher's, exactly as written: condition_encoder, not condition-encoder. A wrong name lists the right ones:

    no component matches "condition-encoder"; repo has:
      CONDITION_ENCODER, LANGUAGE_MODEL, RVQ_DEPTH_DECODER, TRANSFORMER, VOCODER
  • Name the quant (:Q8_0) where a component has several.

  • YuE needs all three of its components before it is recognised.

  • ACE-Step: pull writes a second copy of its text encoder beside the original. Keep both.

Models are found under ~/.dppan/models/bundles at startup. To keep them elsewhere, list the directories:

bash
DPPAN_MUSIC_BUNDLE=/vol/a,/vol/b dppan-orchestrator

Startup logs what it found and why it skipped anything — an incomplete directory names its missing components. A model shown in the dashboard can make a song.

Smaller and faster: --f16 ​

--f16 converts an unquantised component to F16 as it installs: MiniMax's 9.1 GiB transformer becomes half that, and a song takes about half as long. It is lossy, so it is opt-in; already-quantised components are skipped.

To convert a component already on disk, delete it and pull it again with --f16:

bash
rm ~/.dppan/models/bundles/minimax-music3/transformer_bf16.gguf
dppan pull hf:audio-cpp/MiniMax-Music3-GGUF:BF16 --component transformer --f16 \
     --out ~/.dppan/models/bundles/minimax-music3

Do delete the original: two copies of one component leave the choice to chance.

Generating ​

bash
curl http://localhost:8080/v1/audio/music \
  -H 'content-type: application/json' \
  -d '{"model":"minimax-music3","input":"warm lo-fi hip hop, dusty piano, vinyl crackle",
       "lyrics":"[verse]\nlate night on an empty street\n",
       "duration_seconds":30}' --output song.wav
Field
modelThe model, named by its directory
inputThe style: genre, instruments, mood, tempo. Detail helps
lyricsOptional for minimax and acestep; required for yue, with [verse] / [chorus] headers. Leave it out for an instrumental
duration_secondsClamped to the deployment's ceiling
stepsSolver steps. Lower is faster and blurs the words
stemsmix (default), vocal, or instrumental
response_formatwav only
minimax-music3, 20 seconds (the ceiling raised as in Length), from a style of catchy synth-pop at 118 BPM with a clear female lead vocal, and lyrics.

Expect this request to time out. A song takes minutes, so the wait usually ends with a job ID to poll on /v1/generations/{id}. For anything but a short clip, use the job API directly.

Vocals or backing track alone ​

YuE can return the vocals or the instrumental alone at no extra cost:

bash
curl http://localhost:8080/v1/audio/music \
  -H 'content-type: application/json' \
  -d '{"model":"yue-s1-s2-en-cot","input":"slow blues, male vocal","lyrics":"[verse]\n...","stems":"vocal"}'

minimax and acestep make one mixed track and refuse anything but mix. supports_stems in capabilities says which is which.

Progress ​

A job reports which phase it is in. Show the phase name; the fraction moves but its total can shrink.

Phases
minimaxwriting → solving → decoding
acestepsolving → decoding
yuewriting → refining → decoding

Length ​

Time grows with length and steps. On an M-series laptop a ten-second clip takes a minute or two. Songs are capped at ten seconds by default; raise it in frames at 25 Hz:

bash
DPPAN_MUSIC_MAX_FRAMES=1500 dppan-orchestrator   # one minute

Configuration ​

VariableDefault
DPPAN_MUSIC_BUNDLEunsetModel directories, comma-separated. Unset uses ~/.dppan/models/bundles
DPPAN_MUSIC_MAX_FRAMES250 (10 s)Longest song, in frames at 25 Hz
DPPAN_MUSIC_STEPS0 (the model's own)Default solver steps
DPPAN_MUSIC_THREADS0 (the machine's)CPU threads for music
DPPAN_MUSIC_SYNC_TIMEOUT_SECS900How long /v1/audio/music waits before returning a job ID to poll

Default steps by model:

ModelSteps
minimax30Time grows with it
acestep8Fixed; another count is refused
yue—Has no step count

Known limitations ​

  • minimax's default shift and guidance are estimates, as the publisher does not give them. If a song sounds wrong, both can be set per request.
  • acestep without a reference clip has an arbitrary timbre.

Free to run · Proprietary