Music ACE Step

Music (ACE-Step)

ACE-Step task controls in their unloaded state. Visible transformations have separate model requirements. Captured September 25, 2026.

Music is ACE-Step 1.5: DiT-based 48 kHz stereo song generation. The sidebar label is Music; the page heading says ACE Step. It is not MiniMax Music 3 and not Audio Lab.

Open Audio → Music.

Sub-tabs across the top: Generate, Simple, Planner, A2A, Repaint, Edit, Extend, Cover, Extract, Lego, Complete, Vocal→BGM, Lyric→Vocal, Text→Samples, Jobs, Outputs, Install.

Lyric→Vocal and Text→Samples are blocked unless upstream assets exist. They are documented last so you do not plan a session around them.

Live status card

The top card is the worker:

• Worker: not running — click Load…
• Worker: ready plus badges for model, LM, optional VAE swap, attached LoRAs
• CLAP via audio_lab green if Audio Lab has CLAP loaded (ranked generate will score); orange if unavailable (ranked still generates, unranked)
• Change… / Unload all

If runtime dependencies are missing, repair the app venv/setup before any load. If a native model is installed but the shared v1.5 core is missing, use Install shared core. New native installs include the core; old partial installs may not.

Generation forms stay disabled until a base model is loaded. Loading the worker process from Runtime is not enough.

Install (sub-tab)

Install DiT models, LMs, VAE swaps, LoRA packs, and custom HF repos here (or from Model library / omni-cli ace-step install-*).

Practical stacks:

Goal	Model	LM	Notes
Fast general songs	ace-1.5 / 2B turbo	ace-lm-0.6b	Verified through 120 s; turbo defaults ~8 steps, CFG 1.0, shift 3.0
Higher quality on two GPUs	ace-xl-sft	ace-lm-4b on the other GPU	4B LM needs the second card cleared
XL turbo	ace-xl-turbo + 1.7B LM	~19 GB VRAM in tests; cold page-in can take many minutes
Base-only extract/lego/complete	ace-base / ace-xl-base	defaults ~50 steps, CFG 4.0

LoRA packs with verified attach/detach: raspy, acoustic, Chinese New Year, lofi, raga. Start multiplier 0.4–0.5 (safe band about 0.2–0.7). The legacy Chinese Rap package is installed-incompatible (not PEFT) and blocked.

Installs are background jobs. Poll them. Delete is immediate; unload first.

Load…

The load modal/form asks for:

• model variant (required for generation)
• LM variant (required for lyric/caption paths that need the 5 Hz LM)
• VAE default or swap
• device and optional lm_device (ACE can put DiT on the primary GPU and the LM on an auxiliary GPU — component split, not HF layer sharding)
• bf16 (default), optional CPU offload, optional int8 on XL, optional torch.compile (slow first run)

Only send components you want to change. A model hot-swap frees the previous weights first. Do not enable CPU offload on a GPU that already fits.

After load, confirm the live card. Then generate with autospawn conceptually false: the model is already resident.

Generate

Main song form.

Do this:

1. Prompt — style, genre, vocal character, mix notes.
2. Lyrics — optional. Use [Verse], [Chorus], [Bridge] tags. Omit lyrics and say instrumental, no vocals for beds.
3. Negative prompt — on v1.5 this is the LM negative prompt.
4. Duration — 1 to 600 seconds documented. Long tracks should still be composed from sections; listen for continuity.
5. Steps / CFG / scheduler (euler, heun, dpmpp) / shift (1.0–5.0; turbo 3.0, SFT/base 1.0).
6. Optional BPM (30–300), keyscale (D minor), time signature (4 for 4/4), seed, guidance interval.
7. Generate or ranked N=1–16.

The call is synchronous. The returned job_id is an output folder, not a background job to poll. Play the WAV from the result, Jobs, Outputs, or Media library (Omni source).

infer_method: ode is the official default (use for SFT). sde is known to garble SFT. Leave Adaptive Dual Guidance off unless you know you want it.

Ranked generate

n candidates, optional CLAP score prompt. Without Audio Lab CLAP, results are unranked. With CLAP, pick the best_ file and still listen; CLAP is not a mix engineer.

Simple and Planner

• Simple — style query → caption/lyrics/BPM/key then generate in one operator action (create-sample then generate).
• Planner — LM planning helpers (format-sample, understand). Use when you want the 5 Hz LM to rewrite metadata; turn thinking off on Generate if you want to skip CoT rewrite even with an LM loaded.

Language list on the tab is large (en, zh, ja, …). It labels lyric language for the LM; it is not a guarantee of native pronunciation quality.

Transform existing audio

These modes take init audio (upload or a library URL such as an ACE output or a MOSS WAV in the Omni library):

Mode	What it does	Extra fields
A2A	Restyle / denoise toward the prompt	init_noise_level 0–1
Repaint	Replace a time range	mask_start_s, mask_end_s inside duration
Edit	Lyrics-only or remix	only_lyrics or remix; optional source prompt/lyrics
Extend	Prepend or append new audio	prepend/append, extend duration. On SFT this pads silence and repaints the new region
Cover	Re-style a song	describe the target in the prompt
Vocal→BGM	Voice or vocal stem → accompaniment	optional accompaniment prompt. MOSS TTS output URLs are valid sources

Completion and extension are generative, not lossless splices. In current base-model tests the retained portion changed. If you must keep existing audio bit-stable, use composition on library files instead (see Jobs, tokens, composition, and media retrieval).

Analyze (API / advanced) returns BPM, key, loudness. It does not prove the track is musically coherent.

Base-only arrangement

Load ace-base or ace-xl-base first:

• Extract — track_name from the list (woodwinds, brass, fx, synth, strings, percussion, keyboard, guitar, bass, drums, backing_vocals, vocals)
• Lego — rebuild from tracks
• Complete — track_names

These are not a DAW. They need the base model, not turbo-only.

LoRAs at runtime

Load the base model. Attach an installed pack with the exact adapter_file when the pack has several files (required to choose among a declared pack). Detach by name. Status shows recommended multiplier ranges; do not guess filenames inside a repo.

Jobs and Outputs

• Jobs — persisted ACE output jobs, open/delete/zip
• Outputs — files for a job id

Delete removes the folder. ZIP is the Windows-friendly retrieval path along with Media library export.

Lyric→Vocal and Text→Samples (unavailable)

These sub-tabs exist because the ACE-Step product defines the modes.

• Official Lyric2Vocal and Text2Samples weights are unreleased.
• The registry refuses false installs.
• Specialized endpoints are not usable with official weights.
• Do not follow third-party "just download this LoRA" advice unless Omni's status page shows the adapter as installed and enables_mode for that tab. Until then the mode is blocked.

Use Generate with lyrics for sung songs, Vocal→BGM for accompaniment from a voice, and Audio Lab / MOSS-SFX for one-shot samples.

Unload

Unload all on the live card, or unload model / lm / vae separately. Then kill the exact ACE worker on Runtime if you need the GPU for Comfy.

Do not leave XL + 4B LM resident "for later" while you start Music 3.

Related pages

• Audio Lab — CLAP ranking and short beds
• Music 3 — different engine, lyrics required
• MOSS — speech source for Vocal→BGM
• Feature compatibility