Introduction

takakhoo published Audio Sliders on Hugging Face around October 2, 2026—a bank of 35 rank-4 LoRA sliders that continuously steer a frozen ACE-Step 1.5 XL turbo music model while you keep the prompt and seed fixed. Card: takakhoo/audio-sliders.

This is music control, not a new foundation model: same ACE-Step backbone, dialable axes for mood, ensemble, energy, and related attributes. Install path and demo live under the audio-diffusion-control repo linked from the card.

Pro Tip

Lock prompt + seed first, then move one slider family at a time. Audio Sliders are strongest when you A/B a single axis (e.g. energy) across a held clip rather than stacking every LoRA at once—the card is explicit that sliders can leak into neighboring attributes.

What Shipped

FactDetail

Project

Audio Sliders

Author

takakhoo

HF card

takakhoo/audio-sliders (~created 2026-10-02)

Base model

Frozen ACE-Step 1.5 XL turbo

Controls

35 rank-4 LoRA sliders; continuous position

Invariant

Prompt + seed fixed while you dial

Families

text (12 named attrs, e.g. mood / ensemble / energy), real-axes, real-axes-sets, measured-sets, graded

Eval

Measured on 24 held-out prompts; failed sliders excluded

Lineage

Concept Sliders / SliderSpace ideas applied to audio

License

MIT, subject to ACE-Step terms

Demo

Live demo linked from the HF card

Think of the text family as the friendly entry: twelve named attributes you can reason about in plain language. The real-axes / measured / graded families are for builders who want axes closer to measured or set-based control rather than a single adjective dial.

Why Music Makers Care

ACE-Step already gives you fast instrumental generation; what it does not give you out of the box is a clean “more energy / less ensemble / different mood” knob without re-rolling the prompt. Audio Sliders fill that gap with continuous LoRA positions—closer to a mixing desk than to prompt roulette.

The Concept Sliders / SliderSpace lineage matters: this is the image-slider idea ported into a music diffusion/flow setup, with an explicit holdout measurement pass and failed sliders dropped. That is more honest than a vibes-only control pack.

How to Run It

Install the control package (model + demo extras) from the author’s GitHub path:

pip install "audiosliders[model,demo] @ git+https://github.com/takakhoo/audio-diffusion-control"

Then either launch the demo server (linked from the HF card) or drive a Python SliderBank in your own script: load frozen ACE-Step 1.5 XL turbo, attach the LoRAs you care about, fix prompt + seed, and sweep slider positions.

Practical loop:

  1. Open takakhoo/audio-sliders and skim the family list
  2. Start from the live demo for a listen-before-code pass
  3. Reproduce one axis in Python with a fixed seed before you automate batches
  4. Respect ACE-Step terms alongside the MIT wrapper

Limits (Read Before You Ship)

  • No formal listening study yet—measurements are on 24 held-out prompts, not a perceptual panel
  • Sliders leak—moving one axis can nudge neighboring attributes
  • Training focused on ~10 s instrumental clips; do not assume long-form vocal stems behave the same
  • Uses a custom LoRA layout—expect integration work if your ACE-Step fork assumes a different adapter shape

Conclusion

Audio Sliders is a same-week control drop for people already on ACE-Step 1.5 XL turbo: 35 rank-4 LoRAs, continuous dials, MIT + ACE-Step terms, and a live demo on the HF card. Start at takakhoo/audio-sliders, lock prompt and seed, sweep one family, and keep the leak / short-instrumental caveats in your producer notes.

—Anabel ♡