ExpSoft›ACE-Step 1.5›Help

ACE-Step 1.5 User Guide

Describe a style, paste your lyrics and get a finished song on your own PC. Then re-perform it in another style, borrow its vibe for a new track or make it longer, keep everything in a gallery, and write new lyrics with a local model. Five steps take you from the install to your first songs, remixes and lyrics.

ExpSoft ACE-Step 1.5 v1.2.0.0 5 steps · about an hour, most of it downloading Complexity: Medium

Before you start

A demo, a jingle or a backing track made by hand means writing, recording, arranging and mixing — days of work for one minute of music, and the same again for every variation you want to hear. ACE-Step 1.5 turns a line of text and your lyrics into a finished 60-second track in well under a minute on a recent NVIDIA card, so you can hear ten versions of an idea in the time one used to take. Everything runs on your own PC.

  • ACE-Step 1.5 comes with the ExpSoft subscription on Patreon, from The Power User tier (Tier 2) and above.
  • Windows 10 or 11, 64-bit, and the free .NET 8 Desktop Runtime. If it is missing, Windows offers to download it the first time you start the app.
  • An NVIDIA card. 12 GB of video memory or more is recommended for the XL models; with 8 GB, use the smaller models (Step 2).
  • 16 GB of system memory at least. The largest lyrics model needs about 9 GB of it.
  • Disk space: about 27 GB after the setup, about 55 GB with one XL model, about 95 GB with all of them — plus 2 to 9 GB for the lyrics model you choose.
  • A Hugging Face token is optional. The models download without one; add yours in SETTINGS only if a download asks for it.
Tested configuration. ExpSoft tools are tested and validated on PCs with 16 GB of video memory and 64 GB of system memory for the default settings. With more or less memory, the settings described in Memory settings let you adapt — the experimenting happens on your PC.
Everything stays on your PC. The app, its Python, the models and your songs all live inside the app's folder — nothing is installed on your system and no administrator rights are needed. The app goes online to download what the setup and the models need, to check for a new version and to show the More Tools ? tab. Your prompts, lyrics and songs are never uploaded.

The window at a glance

The WELCOME tab of ExpSoft ACE-Step 1.5 with the twelve tabs, the version badge, the Quick Start in four steps and the requirements
The WELCOME tab. Click the image to enlarge it.
  1. 1
    The tabs — Setup installs everything, Generate makes songs, Remix re-imagines a track, Gallery keeps your songs, LoRAs, datasets and lyrics, Lyrics writes lyrics with a local model, Help holds lyric templates, Dataset and LoRA are for experimenters (see below), Disk shows what takes space, Settings holds your keys, More Tools ? shows the other ExpSoft tools.
  2. 2
    Version badge — says whether a newer version is available.
  3. 3
    Quick Start — this guide in four lines.
  4. 4
    Requirements — the card, the memory and the disk space the app needs.

Every button has a tooltip: hover it to read exactly what it does. The bar at the bottom shows the system and video memory in use.

Step 1

Install

Open the SETUP tab and click Install All. It happens once.

The SETUP tab with the Check Prerequisites and Install All buttons, the installation steps ticked green, and the log reporting every component as OK
SETUP, once everything is installed.
  1. 1
    Check Prerequisites — looks at what is already installed and at your card, and ticks what is there. It installs nothing.
  2. 2
    Install All — nine steps in order: a portable Python, its package manager, a portable Git, the card check, the two code bases, the Python libraries, the 2B model (about 9 GB) and the local lyrics engine. A step already done is skipped.
  3. 3
    The steps — a green tick means done; scroll the list to see all nine.
  4. 4
    The log — every command and its result, also saved in logs\.

The setup downloads about 20 GB. On our test PC, with a fast connection and the 2B model already in place, the whole setup took under 3 minutes.

Then download an XL model. In GENERATE, pick a model and click Download this model (about 19 GB each). The first time you generate with one of the v1.5 models, the app also fetches the parts they all share (about 10 GB) — once.
Step 2

Your first song

Open the GENERATE tab, describe the style, paste your lyrics, click GENERATE.

The GENERATE tab: ten quick examples, a synth-pop style description, lyrics in verse and chorus sections, the path of the finished song and the log of the run
A 60-second synth-pop song with lyrics written for ExpSoft.
  1. 1
    Quick examples — ten ready-made styles. One click fills the style, the lyrics and the length; steps and guidance follow the model you picked.
  2. 2
    Description / Style — genre, instruments, voice, mood, tempo, key: upbeat synth-pop, bright analog synths, female lead vocal, catchy chorus, 118 bpm.
  3. 3
    Random from gallery — for a batch, takes a different lyric from your gallery for every track.
  4. 4
    Lyrics — sections in brackets: [verse], [chorus], [bridge]. Leave [instrumental] for a track without singing.
  5. 5
    Last generated file — where the song was saved.
  6. 6
    The log — the progress of the run.
The lower part of GENERATE: the voice helpers, the model choice set to XL SFT, and the duration, steps, guidance, scheduler and seed fields
The model and its settings, lower in the same tab.
  1. 1
    Voice helpers — buttons that add voice tags to the style: gender, timbre, age and style, accent or language.
  2. 2
    Model — see the table below. Steps and guidance are set for you when you switch.
  3. 3
    Duration — in seconds; 30 to 120 is the sweet spot.
  4. 4
    Steps — more is slower and slightly cleaner. 60 for XL (SFT), 50 for XL (Base), 8 for the turbo models.
  5. 5
    CFG scale — how closely the song follows your text: 15 is the balanced default, lower is looser. Turbo models do not use it.
  6. 6
    Scheduler — euler is fast; heun is about twice as slow and slightly smoother on held notes.
  7. 7
    Seed — keep the same seed and settings to get the same song again; leave it empty for a new one each time.

Further down: Use LoRA (for experimenters), the two memory options, the Title of the file and the Count — more than 1 makes a batch, each track with a new seed, for an overnight run. Every song lands in output\ with a .json file that keeps its settings.

ModelWhat it is forSettings
XL (SFT) — defaultThe best quality60 steps, guidance 15
XL (Base)More variety, less polished50 steps, guidance 15
XL (Turbo)Quick drafts: 8 steps instead of 608 steps, no guidance; no Extend
2B (Turbo)Cards with 6 to 8 GB of video memory8 steps, no guidance; no Extend
2B (v1.0 legacy)The older generation; takes LoRAs trained on the 2B stack60 steps, guidance 15; not in Remix
Step 3

Remix a track

Open the REMIX tab, pick a song, choose what to do with it, click REMIX.

The REMIX tab set to Cover: the source song, the mode, the strength slider at 0.70, an acoustic unplugged prompt, and the log of the finished cover
An acoustic cover of the song from Step 2.
  1. 1
    Input audio — drop or pick a WAV, MP3 or FLAC file; lossless is better.
  2. 2
    Mode — Cover, Inspire or Extend (table below).
  3. 3
    Strength — how close to the source: 0.8 to 1.0 for a cover, 0.2 to 0.5 for inspiration. Not used by Extend.
  4. 4
    Prompt — the new style: acoustic unplugged version, fingerpicked acoustic guitar, soft piano. Lyrics are optional: a cover keeps the original ones.
  5. 5
    Last remixed file — where the result was saved.
  6. 6
    The log — the progress of the run.
ModeWhat you get
CoverThe same song — melody and words — performed in the style you describe.
InspireA new song that borrows the groove and energy of the source.
ExtendThe source kept as it is, followed by the number of seconds you set in Duration. XL (SFT) and XL (Base) only.
The memory options of REMIX with the text encoder option ticked, and the first log line saying the option was switched on for a 16 GB card
On a card with 16 GB or less, the app turns the memory options on by itself.
  1. 1
    Memory options — a remix first reads your whole song, which takes extra video memory. On cards with 16 GB or less, Cover and Inspire switch on the text-encoder option, and Extend switches on both.
  2. 2
    The log says so — the first line tells you which option was switched on.
Step 4

The gallery

Every song you make is listed in GALLERY › Musics, with its settings. Nothing is deleted automatically.

The GALLERY tab: six tracks, the selected song with its settings, its waveform, the player buttons, and its style and lyrics
Six tracks made on the test PC: the song, its cover, an extension, an inspiration, an orchestral and a lo-fi piece.
  1. 1
    Sub-tabs — Musics, LoRAs, Browse Datasets, Lyrics.
  2. 2
    Refresh / Folder — reload the list, or open output\.
  3. 3
    Your tracks — newest first. Right-click a track to rename it, copy or cut it (paste it in Windows Explorer), show it or delete it.
  4. 4
    → MP3 — saves an MP3 next to the WAV; the WAV is kept.
  5. 5
    Reveal / Delete — for the selected track.
  6. 6
    Waveform — click to jump, drag to select a region.
  7. 7
    Player — Play, Pause, Stop; Loop region repeats the part you selected.
  8. 8
    Prompt and lyrics — exactly what made this track.
Step 5

Lyrics

Staring at an empty page is the slow part of a song. The LYRICS tab writes a first draft in the right format from a one-line brief, with a model that runs on your processor — the graphics card stays free for the music.

The top of the LYRICS tab: the engine status, the model choice marked Downloaded, the Download and Unload buttons, a brief and the prompt mode
A brief for the song of Step 2.
  1. 1
    Engine status — grey until your first generation starts it, green when ready. Orange means it is stuck: click Restart Ollama next to it.
  2. 2
    Model — three sizes (about 2, 5 and 9 GB of system memory), picked from your memory and remembered. Downloaded tells you it is already on disk.
  3. 3
    Download / Unload — download the model once, or free its memory when you are done.
  4. 4
    Brief — theme, style, mood.
  5. 5
    Prompt mode — Minimal for the small model, Detailed (more section tags and an example) for the larger ones.
The lower part of LYRICS: the Generate Lyrics and Batch buttons, the generated lyrics in verse, chorus, bridge and outro sections, and the history with Load, Copy and Delete
The draft, written in 48 seconds on the processor, and the history.
  1. 1
    Generate Lyrics — writes one draft and saves it to the history.
  2. 2
    Batch — several drafts in a row; the model stays loaded between them.
  3. 3
    The lyrics — edit them here, then copy them into GENERATE.
  4. 4
    History — every draft is kept in output\lyrics\, and in GALLERY › Lyrics where you can tag and filter them.
  5. 5
    Load / Copy / Delete — bring a draft back into the editor, copy it, or remove it.

Prefer an online assistant? The HELP tab has five ready-to-paste templates (pop-rock, rap, ballad…) that ask for lyrics in the same format.

For experimenters: LoRA

The DATASET and LORA tabs can build a training set from your own tracks and train a LoRA. This part is not supported: in our tests, LoRA training does not give usable results yet. The tabs stay in the app for those who want to experiment, with their own guide: Dataset & LoRA. Every song, remix and example of this guide and of the product page was made without a LoRA.

Memory settings

The defaults are set for a card with 16 GB of video memory. With less, or for long songs:

  • Offload VAE + text encoder to CPU — the lighter option. On our 16 GB card it took a 60-second Cover from the full 16 GB down to 13.3 GB, and the Cover was even a little faster, because nothing had to spill over into system memory.
  • Offload DiT to CPU — the big saving, for songs longer than 2 minutes on 12 to 16 GB cards; generation becomes much slower.
  • A lighter model — XL (Turbo), or 2B (Turbo) for 6 to 8 GB cards.
  • A shorter duration — memory grows with the length of the song.

Remix switches the options on by itself on cards with 16 GB or less (Step 3). The lyrics model runs on the processor and uses system memory only.

How fast is it?

Measured on our test PC (RTX 5080, 16 GB of video memory, 64 GB of system memory). Your card will give other numbers.

TaskTime
A 60-second song with lyrics, XL (SFT), 60 steps28 to 41 seconds, model loading included (13.6 GB of video memory at the peak)
A 60-second Cover, XL (SFT)31 seconds (13.3 GB at the peak, text-encoder option on)
Extend a 60-second song by 15 seconds, XL (SFT), both options onDone in one run (11.7 GB at the peak)
Lyrics with the largest model, on the processor48 to 56 seconds, engine start included

Files the app writes

Everything the app installs and writes stays in its folder.

  • output\ — your songs and remixes, each with a .json of its settings; output\lyrics\ — your lyrics.
  • models\ — the music models; runtime\ — the portable Python, Git and the lyrics engine with its models; venv\ — the Python libraries; cache\ — the download cache.
  • datasets\, loras\, training_profiles\ — your datasets, LoRAs and training recipes.
  • logs\ — one log per run; config.json — your settings and keys.

The DISK tab shows what each folder takes, with a Delete button on the ones that are safe to clear.

Troubleshooting

Windows asks me to install .NET when I start the app

The app needs the free .NET 8 Desktop Runtime. Accept the download Windows offers, install it, then start the app again.

"This model is not downloaded yet"

Click Download this model under the model choice. The first song with a v1.5 model also downloads the shared parts once (about 10 GB): that first run takes longer.

Out of memory, or the screen flickers and the driver restarts

The card ran out of video memory. Tick Offload VAE + text encoder, then Offload DiT if needed, shorten the song, or pick a turbo model. Close other programs that use the graphics card.

Extend is refused

The turbo models cannot extend a song. Pick XL (SFT) or XL (Base).

The lyrics engine is grey, or seems stuck

Grey is normal: the engine starts with your first Generate Lyrics. If the status turns orange or a generation never ends, click Restart Ollama. Closing the app stops the engine and frees its memory.

"Smart App Control blocked a Python extension"

A Windows policy of your PC, not a broken install: run the same step once more.

The More Tools tab is blank

It needs the WebView2 runtime and an internet connection.

Something went wrong. What should I send?

The logs\ folder has one file per run (setup, generation, remix, lyrics, training). Send the one of the run that failed to ExpSoft, with a word on what you clicked.