Voxta docs

ACE-Step

Songs with vocals, or instrumentals, made on your own GPU by ACE-Step 1.5, from a style and lyrics.

ACE-Step 1.5 (MIT) makes whole songs, with vocals or instrumental, in a few seconds per song on a recent GPU. It sings in 50+ languages, and can repaint part of a track, extend it, or remix it.

Experimental and admin-only.

Setup

Add the service

Add Modules → Services → ACE-Step → Install, then follow the install steps. Voxta installs ACE-Step into its Python and downloads the turbo model with its VAE and text encoder (~6 GB). It writes 320 kbps mp3 with FFmpeg. On Windows the install downloads it, and on Linux it uses the system's (see FFmpeg). Watch the Terminal for progress.

Pick a preset

PresetMusic modelPlanner
BalancedTurbo0.6B
LightTurboNone
BestTurbo1.7B
DetailedSFT1.7B

A planner or the SFT model is downloaded the first time a preset that uses it loads.

The model starts on the first track, not with the service, so reading its options or opening a chat takes no video memory.

Models

Music Model:

ModelNotes
Turbo (fast)8 steps. The default.
SFT (detailed)50 steps with prompt guidance. Follows the style more closely, with more detail: a few seconds more per track, about the same video memory.

Song Planner, ACE-Step's language model that plans the melody and arrangement before the track is rendered. Bigger plans better.

PlannerVideo memory, in all
None (lightest)~5 GB
0.6B (balanced)~7 GB
1.7B (best)~13 GB
4B (most knowledge)~17 GB

Avoid (a negative style) reaches the planner only, so the playground shows it only when a planner is set.

Settings

Module settings:

SettingNotes
Models DirectoryACE-Step's checkpoints folder, default Music/ACEStep under your models folder. An existing ACE-Step checkpoints folder can be used as it is.
LoRAs DirectoryWhere presets pick their LoRA from. Default Music/ACEStep-LoRAs.

Preset settings:

SettingNotes
Music ModelTurbo or SFT.
Song PlannerNone, 0.6B, 1.7B or 4B.
Let the Planner Rewrite the StyleOff by default. On, the planner rewrites the style from the lyrics: a woman's song can come out a man's, and "no drums" can grow drums.
Quality Steps1–100. Empty uses the model's own: 8 for turbo, 50 for SFT. More is a little cleaner and slower.
Style Adherence (CFG)SFT only. 1–15, default 7. Higher follows the style more strictly.
Writing InstructionsHow the language model writes the style and lyrics for ACE-Step. The default comes from ACE-Step's own guide.
LoRAA LoRA from the LoRAs directory, or none.
LoRA Strength0–1, default 1.

What it takes

Length10–600 s, default 120 s. Empty lets the model choose.
StyleUp to 512 characters.
LyricsUp to 4096 characters.
Vocal language50+ languages, or auto.
BPM, key, beats per barOptional.
TakesUp to 4 per request, made in one pass.
Repaint and extend3–90 s at a time. A range past the end extends the track.
RemixMakes the take again in a new style and lyrics, at a strength.

ACE-Step has no stems.

LoRAs

A preset can load a LoRA, a style or voice ACE-Step was fine-tuned on. It is swapped in per track, so presets with different LoRAs share one loaded model.

Put LoRAs in the LoRAs Directory, each as ACE-Step's trainer or Side-Step saves it:

  • a PEFT folder (adapter_config.json beside its weights), or
  • a LyCORIS LoKr (lokr_weights.safetensors, or a .safetensors file carrying its LoKr config).

Single-file LoRAs converted for ComfyUI do not load.

Writing for ACE-Step

  • Style: one genre, at most one influence. Name the texture of the sound and the voice. Give the tempo in words. BPM and key go in their own fields.
  • Lyrics: section tags ([Verse], [Chorus], [Bridge]...) on their own lines, a blank line between sections. A tag can say how to sing it: [Chorus - anthemic].

See ACE-Step's tutorial and musician's guide.

For a version without Python, in one GGUF file, see audio.cpp.

On this page