ACE-Step 1.5

ACE-Step combines an autoregressive planning language model, a diffusion transformer, and a VAE. It is the model most of Miso is built around, and the only family with editing routes.

Install this one first.

What it does in Miso

Four tasks run on ACE-Step:

Task	Where	Route
ACE-Step 1.5	Create	`text2music`
Repaint a section	Remix	`repaint`
Cover a take	Remix	`cover`
Light cover	Remix	`cover-nofsq`

audio.cpp exposes more routes than these, including complete, lego and extract. Miso does not wire all of them up.

Vocals

Full singing vocals across 50 or more languages. For an instrumental, set the vocal mode to Instrumental and Miso omits the lyrics.

Options

The guided form covers most of these. The names below are what actually reaches the model.

Option	Does
`text`	Style, instrumentation, mood, production
`lyrics`	Plain text. Section tags like `[Verse]` are read
`duration_seconds`	Target length. `-1` lets the planner decide
`num_inference_steps`	Diffusion steps. Default 8 on Turbo
`guidance_scale`	Default 1.0 on Turbo
`bpm`	Locks tempo
`keyscale`	Locks key, for example `C minor`
`timesignature`	For example `4/4`
`language`	Lyric language code, default `en`
`negative_prompt`	What to suppress
`repaint_start`, `repaint_end`	Repaint window in seconds
`repaint_strength`	0 to 1

text is the one ACE-Step field with no server-side default. Miso always sends it, even empty, because omitting it answers HTTP 500.

Memory

The Turbo Q8 package file is 6.19 GB. Resident VRAM peaks at 13.1 GB without memory saving.

Miso always sends ace_step.mem_saver=true, which releases intermediate graph memory and keeps resident VRAM around 6.8 GB. On a card with 16 GB or less this is what makes it run.

Prompting

Keep the text prompt on musical style, instrumentation and energy. Put the actual words into the lyrics field.

A prompt that works:

energetic pop punk with bright distorted guitars, punchy drums,
driving bassline, melodic male vocals

Lyrics with the structure marked:

[Verse]
Walking down the empty street at night
Chasing shadows in the neon light

[Chorus]
We never look back

Routes do not all behave the same way

This is the thing most worth knowing about ACE-Step, and it is easy to get wrong from the upstream manual.

cover follows a prompt closely. repaint barely does, because it rebuilds the section from the music around it. Both carry the same Planner | Not used line in the manual, so that column does not predict behaviour.

Measured: opposite prompts through repaint produced audio 4 to 13 apart on a brightness measure, where the same prompts through text2music came out 1098 apart.

Practical version: if you want the prompt to decide the result, use a cover. If you want to replace a span and keep the rest, use repaint and drive it with lyrics and seeds.