The Wizard's Lyre 0.1.0

The Wizard's Lyre 0.1.0

The first public release of The Wizard's Lyre: a local, human-guided
generative music studio powered by ACE-Step 1.5. Compose from an idea or an
exact plan, keep an immutable history of every take, reshape selected audio,
train local style packs, and export the result without sending music to a
cloud service.

Install

Download wizards-lyre-0.1.0-linux-x86_64.tar.gz or the matching .zip, plus
SHA256SUMS, from the assets below. Verify the archive before extracting it:

sha256sum --ignore-missing --check SHA256SUMS

Then enter the extracted directory and run:

./scripts/lyre bootstrap
./scripts/lyre doctor

The runtime bundles include the compiled web studio, so Node.js is not
required. Setup downloads the pinned ACE-Step source and model weights; Lyre
does not redistribute those large third-party artifacts.

Highlights

• Local text-to-music generation with simple and custom composition plans.
• Fast iterate, detailed polish, CPU-offloaded 4B quality, and
  structural-editing studio_ops model profiles.
• Cover, waveform repaint, track extraction, instrument addition/replacement,
  and arrangement completion without overwriting the source take.
• Fixed-seed reproducibility, A/B comparison, favorites, notes, quality
  scores, lyric timestamps, local ingest, and DAW-ready zip export.
• LoRA style-pack training and application from eight or more local takes.
• Localhost-only FastAPI/React architecture with a durable SQLite queue and a
  dedicated, crash-recoverable GPU worker.

See the 0.1.0 changelog
for the complete public baseline.

Release verification

Hardware validation was completed on 2026-09-05 with:

Component	Verified configuration
GPU	NVIDIA GeForce RTX 4070 Ti SUPER, 16,376 MiB
NVIDIA driver	595.97
Operating system	Ubuntu 26.04 LTS under WSL2 (6.18.33.2-microsoft-standard-WSL2)
ACE-Step	14c0211d5a0653b0f63e27686f4c3f151b4d8629
FFmpeg	n8.1.2-50-g1a748fe2cd-20260905, shared build
Browser	Chromium 140.0.7339.16

The release gates produced these results:

• ./scripts/lyre smoke-gpu generated audible 10-second, 48 kHz audio on
  CUDA.
• ./scripts/lyre live-check --include-lora passed all 14 stages in one run:
  real LM plan fill and persistence, custom generation, byte-identical seeded
  reproduction, polish, CPU-offloaded quality, cover, repaint, upload,
  extract/lego/complete with model swap-back, profile enforcement,
  annotations, safe export, path-jail rejection, and LoRA train/application.
• Style-pack training found all eight FFmpeg-decoded source files, wrote a
  loadable PEFT adapter, and immediately generated a 15-second, 48 kHz take
  with it in the same worker process.
• Worker recovery was exercised by terminating the worker during generation.
  The server remained responsive; after the 60-second stale-heartbeat window,
  a new worker reclaimed the same job as attempt 2 and completed it with an
  audible 10-second, 48 kHz take.
• The production SPA passed a real Chromium flow covering project creation,
  two real generations, playback, waveform drag-selection, A/B comparison and
  swap, export download, confirmed deletion, and 404 readback.
• The locked Python and web suites, linters, type checks, license compliance,
  dependency audits, source distributions, and reproducible runtime archives
  are repeated by the tag workflow before it creates the draft release.

Verify provenance

SHA256SUMS covers both runtime archives. GitHub also publishes build
provenance for the archive digests; with the GitHub CLI installed:

gh attestation verify wizards-lyre-0.1.0-linux-x86_64.tar.gz \
  --repo wizards-ecosystem/wizards-lyre

Supported boundary

• Linux x86-64 or Windows 11 through WSL2, with an NVIDIA CUDA GPU. The tuned
  baseline is 16 GB VRAM.
• Lyre binds only to 127.0.0.1 and intentionally has no authentication. Do
  not expose it through a public bind, port forward, or reverse proxy.
• FFmpeg shared libraries are required for LoRA style-pack training. Ordinary
  generation remains available without FFmpeg.
• ACE-Step and its model artifacts retain their upstream licenses and notices.
  Review THIRD_PARTY_NOTICES.md before redistribution.
• You are responsible for rights to uploaded audio and lyrics and for
  evaluating generated music before publication or commercial use.

For setup details, model footprints, WSL2 guidance, updates, and backups, see
the installation guide.
For help, start with the
troubleshooting guide
and support policy.