ACE-Step v1.5 Turbo

Most text-to-music models are vibes-only: you say "dark cinematic electronic" and pray. ACE-Step v1.5 Turbo is the one that actually lets you talk about music - BPM, key and scale, time signature, vocal language - and it repaints sections of an existing track too. It's ByteDance's step-distilled audio model (yes, the same lab behind SDXL-Lightning), served by Runware as a cloud node, which means the tradeoff you'd normally make on your own GPU - run the big slow model or the distilled one - is a dropdown instead of a hardware decision.

The "Turbo" in the name is doing real work here. Step distillation collapses the denoising trajectory so the model generates in far fewer steps; the node's steps default is 10, and the ceiling is 20. Same logic that made SDXL-Lightning and Hyper-SD fast locally, just applied to audio and hosted off-box. You pay per generation instead of per watt, and for short tracks that's fractions of a cent.

How it works

Prompt in, song out, but with actual musical parameters riding alongside. settings.bpm (30–300), settings.keyScale in {Note}{Accidental} {Mode} form like C major or F# minor, and settings.timeSignature for beats per measure. settings.vocalLanguage takes an ISO 639-1 code, or unknown for instrumental/auto-detect. Leave any of them off and the model decides - which is the right default for a first take, and the settings you reach for when a take is close but in the wrong key.

The genuinely unusual feature is repainting: feed an input audio via the audio socket (UUID or URL), then set settings.repaintingStart and settings.repaintingEnd to rewrite just a region. Negative start values prepend audio before the region, and end values past the track's duration append new audio. That's a targeted edit tool - swap a bridge, extend an outro - which almost nothing else in this pack's audio lineup offers.

The inputs that matter

• positivePrompt (required) - style and content. Be specific about genre and mood; the model has real parameters to hang your description on.
• steps - 10 by default; drop toward 1 for rough drafts, raise for final takes. Distilled models stop improving fast past a point, so don't grind to 20 thinking you're getting quality.
• duration - 30 to 300 seconds. Pick the length you actually need.
• settings.lyrics - full lyrics in standard song form; pairs with vocalLanguage.
• settings.keyScale / settings.bpm / settings.timeSignature - the music-theory controls above.
• audio + repaintingStart/repaintingEnd - the edit-region path. strength (off by default) controls how much of the source bleeds through when you enable it.

Everything else follows the pack's house style: numberResults for variations, outputFormat for MP3/WAV/FLAC/OGG, and an audioSettings socket for sample rate/channels/bitrate. The single audio (AUDIO) output goes straight into SaveAudio.

Where people stumble

• The repaint region needs input audio - the tooltips say it, and it's the rule people forget: repaintingStart and repaintingEnd are ignored without the audio input connected.
• KeyScale format is strict. F# minor works; F sharp minor probably won't. Match the {Note}{Accidental} {Mode} template.
• The toggle pattern - settings.bpm, settings.repaintingStart, and friends are "enable to set X" booleans with a paired _value field. On = your value wins; off = model default. If your BPM setting seems to do nothing, the toggle is off.
• Distilled expectations - 10 steps is the sweet spot, not a low-quality compromise. This is the "Q8 is basically fp16" of the audio world: the whole point of the Turbo model is that you don't need 50 steps.

It's the most "producer-friendly" music node in the pack - the one you'd reach for when you need a track to actually be in E minor at 128 BPM, not just "sort of electronic."