ACE STEP-1.5 Audio Models - The Essentials

Last updated: July 11, 2026

Covers ACE-Step 1.5 Text to Music - Turbo, ACE-Step 1.5 Text to Music - Quality, ACE-Step 1.5 Cover - Turbo, ACE-Step 1.5 Cover - Quality, ACE-Step 1.5 Repaint - Turbo, and ACE-Step 1.5 Repaint - Quality

ACE-Step 1.5 is an open-source music family that writes, restyles, and edits complete songs, with lyrics you script line by line. Three tasks each come in two checkpoints: Turbo renders fast at a lower cost per generation, and Quality is the full-size checkpoint for final takes.


Which Model Should I Use?

Model

ID

What it does

Text to Music - Turbo

model_ace-step-1-5-turbo-text-to-music

Full songs from a prompt and lyrics, fast

Text to Music - Quality

model_ace-step-1-5-quality-text-to-music

Same task, highest fidelity

Cover - Turbo

model_ace-step-1-5-turbo-cover

Restyle an existing track, fast

Cover - Quality

model_ace-step-1-5-quality-cover

Same task, highest fidelity

Repaint - Turbo

model_ace-step-1-5-turbo-repaint

Regenerate one section of a track, fast

Repaint - Quality

model_ace-step-1-5-quality-repaint

Same task, highest fidelity

Rule of thumb: draft on Turbo, finish on Quality. The two checkpoints take identical prompts and settings, so you can move a draft up without rewriting anything. Cover and Repaint need a source audio asset; they can't generate from scratch.


Text to Music

Text to Music turns two fields into a finished song: a prompt that describes the sound and a lyrics script that says what gets sung, and when. Either field works alone; together they give you the most control.

Build the prompt from genre, mood, two or three named instruments, timbre, era, production style, and vocal character. Set tempo, key, and length with the dedicated bpm, keyscale, and duration controls, never inside the prompt. Tracks run from 10 seconds to 10 minutes.

The lyrics field is a script, not only words:

  • Structure tags on their own lines: [Verse], [Chorus], [Bridge], [Outro], [Fade Out]. Add at most one modifier per tag, like [Bridge - whispered]. Stacked modifiers can get sung as words.

  • (Parentheses) mark backing vocals. UPPERCASE lines get belted.

  • [Instrumental] alone, or the Instrumental toggle, produces music with no vocals.

  • Keep sung lines at 6 to 10 syllables, leave a blank line between sections, and close with [Outro] or [Fade Out] to avoid abrupt endings.

Set Vocal language to any of 50+ languages, or leave it on Auto. Here's the setup behind the ballad below:

prompt: female vocal, piano ballad, emotional, intimate atmosphere,
        strings, building to powerful chorus, warm analog production
lyrics: [Verse 1]
        I fold the morning into paper boats
        Send them down the river one by one
        ...
        [Chorus - powerful belting]
        LET IT GO, let the river hold it
        ...
duration: 170   bpm: 72   keyscale: C Major

Paper Boats: the 170s piano ballad from the prompt above (Text to Music - Quality).

Paper Boats, Text to Music - Quality · Open on Scenario

Byeolbit: a K-pop track sung in Korean, Vocal language set to ko (Text to Music - Quality).

Byeolbit, Korean vocals, Text to Music - Quality · Open on Scenario

Lo-fi hip hop study beat with the Instrumental toggle on (Text to Music - Quality).

Lo-fi instrumental, Text to Music - Quality · Open on Scenario


Cover

Cover re-records an existing song in a new style while keeping its structure recognizable. Pick any audio asset as the Source audio and describe the target sound in the prompt. Then set Cover strength:

  • 0.7 to 1.0 stays faithful to the original arrangement.

  • 0.5 to 0.6 transforms the genre while the hook survives. This is the sweet spot for style flips.

  • Around 0.2 only borrows the mood and rebuilds the rest.

Pass the original lyrics to keep the words through a genre change, or pass new ones to rewrite them. Swapping in a translated lyric sheet with a matching Vocal language localizes a song while its melody carries over. The Instrumental toggle re-performs the track without vocals; it's a regeneration, not stem separation. An optional Reference audio transfers the feel of another track without copying its structure.

srcAudio: your funk disco track
prompt: full symphonic orchestra, acoustic instruments only, cinematic
        arrangement, soaring string section, bold brass, timpani rolls,
        concert hall reverb, dramatic and majestic
audioCoverStrength: 0.55   instrumental: true

The source: a 70s funk disco track.

The cover source: 70s funk disco · Open on Scenario

The cover: the same song as a full orchestra, strength 0.55 (Cover - Quality).

Orchestra cover at strength 0.55, Cover - Quality · Open on Scenario

Localization: the Paper Boats ballad re-sung in French at strength 0.7 with a translated lyric sheet (Cover - Quality).

French localization, Cover - Quality · Open on Scenario


Repaint

Repaint regenerates one time window of a track and leaves everything outside it untouched. Set Start and End in seconds, describe the new music, and the section rebuilds in context. Total duration never changes, and -1 as the end value regenerates through to the end of the track.

Use it to swap a chorus, replace a solo, rework an intro or ending, change a section's energy, or re-render a few flubbed seconds without touching the rest.

To change the sung lyrics inside the window:

  1. Start the prompt with an instruction, like Repaint the selected section with new sung lyrics: followed by your style description.

  2. Pass the full lyric sheet with the new section swapped in place.

  3. Turn Thinking off.

srcAudio: the Paper Boats ballad
repaintingStart: 84   repaintingEnd: 116   thinking: false
prompt: Repaint the selected section with new sung lyrics: female vocal,
        piano ballad, emotional, building to powerful chorus
lyrics: (full sheet, with the final chorus rewritten as "LET IT RISE...")

Before: the original ballad, chorus sings "Let it go" from 1:24.

The repaint source: original chorus · Open on Scenario

After: only 1:24 to 1:56 regenerated, now singing "Let it rise"; everything else is untouched (Repaint - Quality).

Repainted chorus, Repaint - Quality · Open on Scenario

Locate the region on the actual track you're editing. Two takes of the same song place their sections at different timestamps, so a region copied from another take will miss.


Parameters

Parameter

Models

What it does

prompt

all

Style description, up to 512 characters

lyrics

all

Lyric script with structure tags, up to 4096 characters

instrumental

all

Music with no vocals

vocalLanguage

all

Vocal language, 50+ options, Auto by default

duration

Text to Music

Track length, 10 to 600 seconds

bpm, keyscale

Text to Music

Tempo and key anchors

srcAudio

Cover, Repaint

The track to restyle or edit

referenceAudio

Cover

Optional style-transfer track

audioCoverStrength

Cover

How closely the result follows the source, 0 to 1

repaintingStart, repaintingEnd

Repaint

The window to regenerate, in seconds, -1 for end

audioFormat

Repaint

Output file type: mp3, wav, or flac

numOutputs

all

1 to 4 variations per run

thinking

all

Plans the track first; keep it off for repaint lyric swaps

seed

all

Repeatable starting point for the first output

Tip: request wav output to keep a lossless original alongside the streaming copy.


Use Cases

  • Games: boss themes, level loops, victory stingers, and section-level iteration on a locked track with Repaint.

  • Marketing: jingles with sung copy, genre-flipped variations of one brand track with Cover.

  • Film and video: trailer builds, endings reworked to picture with Repaint.

  • Localization: one song re-sung in French, Korean, or Japanese with translated lyrics on Cover.

  • Music production: sketch a song with lyrics, then fix one line or swap one solo without regenerating the track.


Tips for Better Results

  1. Keep tempo and key out of the prompt. Use bpm and keyscale; a tempo written in the prompt fights the controls.

  2. One modifier per structure tag. [Chorus - anthemic] works; a stack of modifiers may be sung.

  3. Make the prompt and lyrics agree. A prompt naming a violin solo with [Guitar Solo] in the lyrics degrades both.

  4. Start covers at 0.55. Raise the strength if the result strays too far, lower it if it hugs the original.

  5. Batch with numOutputs. Takes vary, so generate 2 to 4 and pick.

  6. Verify by listening. Scenario's automatic audio captions are unreliable on music; trust your ears, not the caption.


Known Limitations

  • Outputs vary between runs, even with identical settings and the same seed.

  • A vocal take can occasionally drop its lyrics partway through a long track. Regenerate the take.

  • The prompt's vocal gender isn't always honored. Rerolling or strengthening the vocal descriptors helps.

  • Repaint can't make a track longer; a region ending at -1 regenerates the ending at the same length.

  • Very long tracks can lose structure. The sweet spots are 30 to 60 seconds and 2 to 4 minutes.


FAQ

Should I use Turbo or Quality?

Draft on Turbo, finish on Quality. Prompts and settings carry over unchanged.

Can it sing in my language?

Yes, most likely. The Vocal language control lists more than 50 languages, and Auto detects the language from your lyrics.

Can Repaint extend my track?

No. Repaint always preserves the total duration; it regenerates inside the track, including its ending.


Related

Open the models: ACE-Step 1.5 Text to Music - Turbo · ACE-Step 1.5 Text to Music - Quality · ACE-Step 1.5 Cover - Turbo · ACE-Step 1.5 Cover - Quality · ACE-Step 1.5 Repaint - Turbo · ACE-Step 1.5 Repaint - Quality

Questions? Reach us through the in-app chat.