music-2.5

Use when the user wants AI song generation with vocals, sung lyrics, original tracks from a style prompt, or source audio for a music video.

Music 2.5 (MiniMax · Replicate)

Full-length songs with natural vocals from lyrics + style description. Not a Pruna P-model — runs on Replicate.

Primary workflow: ai-music-video — lyrics → song → lyric-safe cuts → p-video-avatar / p-video clips → assembly.

When to use

Goal	Use this
Sung track for a music video	Yes — write lyrics with section tags first
Drive p-video clip length	Yes — export MP3 → upload to Pruna → audio input
Documentary narration	No — use gemini-3.1-flash-tts
Instrumental bed under VO	No — use stable-audio-2.5

Environment

bash

export REPLICATE_API_TOKEN=r8_...

Requires ffmpeg / ffprobe for slicing and assembly in the music-video workflow.

Model input (Replicate)

Field	Notes
lyrics	Required. 1–3,500 characters. Use structure tags and \n line breaks.
prompt	Optional style string — genre, mood, tempo, vocal timbre, key instruments (up to ~2,000 chars).
sample_rate	16000 · 24000 · 32000 · 44100 (default)
bitrate	32000 · 64000 · 128000 · 256000 (default)
audio_format	mp3 (default) · wav · pcm

Output: audio file URL (typically MP3, ~2:30–4:30 for full songs).

Structure tags

Control arrangement with tags on their own lines (see Music 2.5 readme):

[Intro] · [Verse] · [Pre Chorus] · [Chorus] · [Hook] · [Bridge] · [Solo] · [Inst] · [Build Up] · [Drop] · [Interlude] · [Break] · [Transition] · [Outro]

• One tag per section; follow with 2–4 lyric lines per section for clean melodies.
• \n = line break (also a safe video cut boundary in the music-video workflow).
• \n\n = pause between sections.
• Parentheticals work for ad-libs and directions: (Ooh, yeah) · (Guitar solo — slow, bluesy).

Prompt tips

Follow the model’s prompt guide: genre + mood + vocal description + tempo + instruments + production feel.

Example prompt:

text

Indie pop, uplifting, warm female vocal, 92 BPM, acoustic guitar and soft synth pads,
wide soundstage, crisp modern production, anthemic chorus

Instrumental sections: use [Inst] or [Solo] tags with parenthetical instrument directions instead of sung lines.

HTTP (curl)

bash

curl -s -X POST \
  -H "Authorization: Bearer ${REPLICATE_API_TOKEN}" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "lyrics": "[Verse]\nWe built it line by line\nEvery skill a stepping stone\n\n[Chorus]\nRun the pipeline, watch it grow\nPruna models, let them flow",
      "prompt": "Indie pop, uplifting, warm female vocal, 92 BPM, acoustic guitar and mellow synth pads, no harsh distortion",
      "sample_rate": 44100,
      "bitrate": 256000,
      "audio_format": "mp3"
    }
  }' \
  "https://api.replicate.com/v1/models/minimax/music-2.5/predictions"

Poll urls.get until status is succeeded; download output.

Repo helper

bash

python3 catalog/workflows/verticals/music-video/scripts/generate_song.py \
  --plan output/my-music-video/music_video_plan.json \
  --out-dir output/my-music-video

Good to know

• English and Mandarin have strongest pronunciation; other languages vary.
• Each generation is unique — same lyrics + prompt produce different arrangements.
• Max ~5 minutes per generation.
• Data is sent to MiniMax via Replicate — see their privacy policy.

Related

• audio-post-production.md — when to use songs vs narration vs beds
• ai-music-video workflow
• gemini-3.1-flash-tts — spoken narration (not song)
• stable-audio-2.5 — instrumental beds only
• replicate-api.md