ace-step-xl

Adapt score or song cues into ACE-Step 1.5 XL inputs using task type, caption, lyrics or instrumental structure, BPM, key, time signature, and temporal performance directions.

SKILL.md

GitHub

ACE-Step 1.5 XL Prompting

ACE-Step XL uses the same control pattern as ACE-Step 1.5 with a larger 4B DiT decoder.

Core fields

  • task_type: usually text2music, or another supported transform task when the source cue requires it
  • caption: overall genre, instruments, mood, atmosphere, timbre, vocal identity, production, and progression
  • lyrics: lyric text plus section and temporal performance information. For instrumental score, use [Instrumental] and structure markers as needed
  • bpm: 30 to 300 when a fixed tempo matters
  • keyscale: musical key when useful
  • timesignature: meter when useful

Workflow

  1. Read music cue and score motif rules.
  2. Put global musical identity in caption.
  3. Put time-varying structure, section changes, performance shifts, and lyrics in lyrics.
  4. Do not overload caption with a second-by-second timeline that belongs in structure.
  5. Choose XL variant only as execution metadata: xl-turbo for fast high-quality iteration, xl-sft for tunable higher-quality sampling, xl-base when the task requires base-model transform features.
  6. Save under 04_generation/prompts/ace-step-xl/<cue-id>.json.

Done

Global musical identity and temporal structure are separated cleanly, and the prompt reflects the cue's dramatic job and hit points.