ExpSoft›ACE-Step 1.5›Help›Dataset & LoRA

Dataset & LoRA

A LoRA is a small add-on file that teaches the model the sound of your own tracks. You call it up with a trigger word in your prompt. Three steps: build the dataset, train, use it in GENERATE.

ExpSoft ACE-Step 1.5 v1.2.0.0 Experimental · not supported Complexity: Advanced

Before you start

Getting a model to sound like your music usually means stitching scripts together, labelling every file by hand and converting it to a format the trainer accepts. The DATASET tab reads your folder, keeps one caption per track and writes the training set; the LORA tab trains on it; GENERATE loads the result.

Not supported — for experimenters. In our tests, LoRA training does not give usable results yet. These tabs stay in the app for those who want to try; the songs ACE-Step 1.5 makes need no LoRA.
Experimental. Training takes time and results vary from one dataset to the next: a short training changes the sound only a little, and the right number of passes depends on your tracks. Treat it as a workshop, not a button.
  • Your own music, or music you have the rights to. A LoRA is not a permission to copy the voice or style of an existing artist.
  • 10 to 30 tracks of one coherent sound — one genre, one mood. WAV or FLAC is better than MP3.
  • 16 GB of video memory. On our 16 GB test card the training used all of it and spilled over into system memory: it works, more slowly. Close other programs that use the graphics card.
  • 32 GB of system memory or more.
Step 1

Build the dataset

Put your tracks in one folder and open the DATASET tab.

The DATASET tab: the input folder, the analysis options, the caption and export buttons, a table of twelve lullaby tracks at 70 BPM, and the caption of the selected track
Twelve instrumental lullabies made with the app, one caption each.
  1. 1
    Input folder — Browse… to your folder, then Scan: every audio file is listed, with the captions already saved next to it.
  2. 2
    What to analyse — look up lyrics, detect the language, measure tempo and key, transcribe missing lyrics, write captions. Untick what you do not need (lyrics lookup for instrumentals, for instance).
  3. 3
    Auto-fill · Transcribe · Generate captions — fill the table for you. Captions come from a local model; an optional cloud caption service works with your own free key, set in SETTINGS.
  4. 4
    Export sidecars → Build HF dataset → Train LoRA with this dataset — save one caption and one lyrics file per track, build the training set, and open the LORA tab with it.
  5. 5
    The table — one row per track; tick the ones to use.
  6. 6
    Edit selected row — the caption matters most: describe the sound the way you will write your prompts (solo kalimba and music box lullaby, soft felt piano, slow 70 bpm).
Step 2

Train

In the LORA tab, check the settings and click START TRAINING.

The LORA tab: dataset and output folders, the XL training stack, the xl-base checkpoint, a LoRA adapter, 60 seconds of audio, 100 epochs and the trigger word expsoftlullaby
A training on the lullaby dataset, trigger word expsoftlullaby.
  1. 1
    Dataset — filled in by Train LoRA with this dataset, or pick it with ….
  2. 2
    Output — where the LoRA is written, in loras\ by default.
  3. 3
    Training stack — XL (v1.5): the LoRA will work on the XL models. The 2B stack is kept for the older model.
  4. 4
    Base checkpoint · Adapter type — acestep-v15-xl-base and lora are the recommended pair.
  5. 5
    Audio cap · Epochs — seconds kept per track (60 fits a 16 GB card), and how many times the training goes through the whole dataset. Small datasets need many passes.
  6. 6
    Trigger word — a made-up word added to every caption; you will type it in your prompts.
  7. 7
    Progress — see below.
A training in progress: step 20, 50 percent, elapsed time, time left and loss, the checkpoint buttons, and the training log
A two-pass test run, half-way.
  1. 1
    Progress — phase and step, percentage, elapsed time, time left and loss.
  2. 2
    Latest LoRA checkpoint — open the LoRA folder or the output folder.
  3. 3
    The log — a checkpoint is saved at the end of each pass.
  4. 4
    START TRAINING — greyed out while a training runs.

On our test PC (RTX 5080, 16 GB), twelve 60-second tracks were prepared in 18 seconds, then each pass through the dataset took about 45 seconds. The training is tuned for the base and SFT models, which are used with guidance when you generate.

Step 3

Use your LoRA

Every training run appears in GALLERY › LoRAs.

GALLERY, LoRAs sub-tab: one LoRA card with its trigger word, rank and size, and its sidecar with trigger token, base model, dataset and notes
The LoRA of the test run, with its notes.
  1. 1
    Your LoRA — one card per training run. Right-click it and choose Use in Generate: GENERATE opens with Use LoRA ticked and the path of this LoRA filled in.
  2. 2
    Trigger token — the word to put in your prompt.
  3. 3
    Base model — the model it was trained on.
  4. 4
    Notes — the training settings, written by the app; add your own.
  5. 5
    Save sidecar — keeps your edits.

In GENERATE, start your style with the trigger word — expsoftlullaby, soft piano, rain — and set LoRA scale: 0.5 is a hint, 0.8 the default, 1.0 the full effect.

Tips

  • Coherence beats quantity: 15 tracks with one sound train better than 40 mixed ones.
  • One caption style for the whole dataset, close to the prompts you will write.
  • A unique trigger word, one that means nothing else.
  • Compare: generate the same prompt and seed with and without the LoRA, and at different scales.
  • Profiles: the LORA tab can save a set of settings with your notes, and comes with two reference recipes for the 2B stack.

Troubleshooting

The training is very slow

Close other programs that use the graphics card, lower Audio cap (30 to 45 seconds), and keep Batch size at 1.

"No labelled sample"

The tracks have no captions yet. In DATASET, write or generate a caption for each track, then click Export sidecars.

The LoRA seems to change nothing

Check that the trigger word is in your prompt and that Use LoRA is ticked, raise the scale, or train for more passes — short trainings change the sound only slightly.

A LoRA is refused when I generate

A LoRA only works on the model family it was trained on: an XL LoRA on the XL models, a 2B LoRA on the 2B (v1.0 legacy) model.

← Back to the ACE-Step 1.5 User Guide