Turn a prompt into a finished-sounding song.
A standalone desktop music studio for the open-source ACE-Step model —
describe the vibe, write (or auto-generate) lyrics, audition takes, and drag
the result into any DAW. Optional Ableton Live link for tempo/key sync and
clip drop-in.
Product & creative direction by Alec Lewis. Application developed with AI coding assistance.
The music generation itself is ACE-Step's
open-source model. Backline is the studio built around it — everything in
ableton-sidebar/:
- A composer, not a form — style caption, structured BPM / Key / Time /
Duration / Vocal language fields (any of them AUTO), and lyrics with
[verse]/[chorus]structure tags, sent to the engine as real parameters. - An audition workflow — Spotify-style results list, hover-to-play, waveform scrubber, variation groups with a Compare view, persistent history.
- Reliability as a feature — job queue with live progress, diagnostics panel, and a smart Fix it button that revives a crashed engine, reloads models, and retries the failed job.
- A path into your DAW — download the WAV, native drag-out to anywhere, or a Max for Live bridge that syncs Live's tempo/key and inserts clips straight into a Session slot.
- The desktop shell — Electron window with tray and always-on-top pin, a zero-dependency Node sidecar serving the React UI.
📖 Full Backline documentation → — features, architecture, Ableton Live setup, troubleshooting, and honest known limits. Design narrative in JOURNAL.md.
You need: an NVIDIA GPU (8 GB works — the launcher applies low-VRAM-safe settings), Python (installed by the engine setup below), and Node 18+.
- Set up the ACE-Step engine — run
install_uv.bat(Windows) orinstall_uv.sh. Models download on first run. Engine details are in the ACE-Step repo. - Build the Backline UI (once, and after UI changes):
cd ableton-sidebar/ui && npm install && npm run build - Launch —
start_ableton_sidebar.batstarts the engine (port 8001) and the Backline sidecar (port 8765) with GPU-safe settings (ACESTEP_DTYPE=float32,ACESTEP_OFFLOAD_DIT_TO_CPU=true,ACESTEP_QUANTIZATION=auto) already set. The first generation triggers a one-time model load + quantization — how long it takes depends on your hardware; the UI shows "loading models" and waits automatically. - Open http://127.0.0.1:8765, type a prompt, hit Generate — or run the
desktop app:
cd ableton-sidebar/shell && npm install && npm start.
The optional chat assistant is your own AI: add your own API key in
Settings (Anthropic / OpenAI / OpenRouter / any OpenAI-compatible endpoint)
and that model assists you in-app. No key ships with Backline — yours stays
in ~/.ace-sidebar/config.json on your machine. Generation needs no key.
To be clear about credit: this repo is a fork of
ACE-Step/ACE-Step-1.5 — the model,
the acestep/ engine code, and the platform launcher scripts are theirs;
see their repo for the model's own documentation.
My work is Backline: the entire ableton-sidebar/
directory (Electron shell, Node sidecar, React UI, Max for Live device),
start_ableton_sidebar.bat, and small patches to the engine's init service
so it runs well on 8 GB GPUs.
MIT — see LICENSE. ACE-Step is by the ACE Studio & StepFun team; Backline is an independent application built on it.