Audio Models - Text to Music (ACE-Step 1.5)

Audio Models - Text to Music (ACE-Step 1.5)

🛠

Text to Music — 48 kHz Stereo, 10 Seconds to 10 Minutes

ACE-Step 1.5 turns a written description of a style into a finished piece of music. Add lyrics and it sings them; leave them out and you get an instrumental. Output is 48 kHz stereo, anywhere from 10 to 600 seconds — long enough for a full track, not just a loop.

CNAPS Studio runs the XL Turbo checkpoint: a guidance-distilled flow-matching DiT that generates in 8 denoising steps. Weights are MIT; the text encoder it ships with is Apache 2.0

What it does

Two text inputs, one audio output. The first describes the music — genre, instrumentation, tempo, mood, production character. The second is optional: lyrics, with section tags like [Verse] and [Chorus] to mark structure. The model returns a stereo track at CD-beating sample rate.

The length range is the part worth pausing on. Ten minutes of coherent generated music is a different proposition from the 30-second clips most open music models produce — it's the difference between a sample and a track. The model is built for it: the released pipeline recommends VAE tiling for long-form generation specifically to keep decode memory bounded, which is a design accommodation for output lengths that would otherwise not fit.

"Turbo" refers to guidance distillation. The published pipeline defaults to 8 denoising steps rather than the dozens a non-distilled diffusion model would need, and the speed comes from that rather than from a smaller or weaker model.

Problem it solves

  • Background music you actually own the pipeline for — Scoring a product video, a demo reel or a course module without licensing a stock track or waiting on a composer
  • Lyrics that get sung, not spoken — This is a music model, not a TTS model with a backing track. The vocal is generated as part of the piece
  • Instrumental by omission — Leaving lyrics empty produces an instrumental. No separate mode to select
  • Full-length output — Up to 600 seconds in one generation, rather than generating a loop and tiling it
  • Eight steps — Guidance distillation makes iteration practical. Generating five variations to pick from is a reasonable thing to do
  • Permissive weights — MIT, with no territorial or revenue conditions

Input/Output

  • Input: two text ports. In CNAPS Studio a filled connection point means required, a hollow one means optional
    • Style prompt (required) — what the music should be. Genre, instruments, tempo, mood, production character
    • Lyrics (optional) — words to be sung, with section tags such as [Verse] / [Chorus]. Leave it unconnected for an instrumental
  • Options:
    • Style prompt (required) — what the music should be. Genre, instruments, tempo, mood, production character
    • Lyrics (optional) — words to be sung, with section tags such as [Verse] / [Chorus]. Leave it unconnected for an instrumental
    • Advanced Options:
      • Diffusion steps (default 8) — the node's own help text notes the turbo checkpoint is distilled for 8, which matches the published pipeline default. Raising it is not the quality dial it would be on a non-distilled model
      • How closely to follow the style prompt (default 7) — see the note below before using this
      • Random seed (1 draws a new one each run) — fix it to reproduce a take
  • Output: a single audio stream — 48 kHz stereo. Port identity is inferred from port order and colour; confirm against the node before relying on it in a tutorial.

Performance & Deployment

Checkpoint
ACE-Step v1.5 XL Turbo (Diffusers format), guidance-distilled
Architecture
Flow-matching DiT
Parameters
The model card describes a 5B DiT (hidden_size=2560, 32 layers, 32 heads; condition encoder encoder_hidden_size=2048). The repository's own parameter readout reports 4B. The discrepancy is not explained upstream — treat "roughly 4 – 5B" as the honest figure
Precision
BF16
Denoising steps
8 (pipeline default; the checkpoint is distilled for it)
Sampling
FlowMatchEulerDiscreteScheduler, recommended shift=3.0
Output
48 kHz stereo
Duration
10 – 600 s per the node description
Long-form handling
The published pipeline recommends enabling VAE tiling for long-form audio to keep decode memory bounded
End-to-end node timing
Not yet measured — verify before committing to a batch schedule
‣
Technical Details (for engineers)

Model Source

  • Hugging Face:

Compliance & Provenance

Provider
ACE-Step
Provider type
Specialized
License
Model weights MIT; bundled text encoder Apache 2.0. Both permit commercial use with no territorial or revenue conditions. Permissive licensing covers the weights — it is not a statement about training-data provenance, and it is not IP indemnification
EU AI Act risk class
Limited Risk
Art. 50 transparency
Required — output is AI-generated audio and is marked with a visible AI-generation label and machine-readable provenance. See AI Policy §2.
Region availability
Available globally
Training data summary
Not published as a structured summary per Art. 53(d)

For more on how we classify models and mark outputs, see our AI Policy.