Offline AI music generation suite. Generate songs, compose MIDI, synthesize vocals, separate stems, create SFX, and master tracks — all locally on your machine.
git clone https://github.com/SysAdminDoc/SlunderStudio.git
cd SlunderStudio
py -3.12 -m pip install -r requirements.txt
py -3.12 main.pyPython 3.11 or 3.12 is required. Install core dependencies explicitly before launch; if anything is missing, Slunder Studio opens a diagnostics screen with the exact setup command. AI models are downloaded on-demand from HuggingFace via the built-in Model Hub.
The same engine, durable-job, export, provenance, and error contracts are available without starting the Qt shell:
py -3.12 slunder_cli.py lyrics "write a dark synth-pop chorus" --output chorus.txt
py -3.12 slunder_cli.py midi "syncopated piano groove" --output groove.mid --demo
py -3.12 slunder_cli.py sfx "rain on a window" --output-dir renders --demo
py -3.12 slunder_cli.py export source.wav renders/master.wav --format wav
py -3.12 slunder_cli.py provenance renders/master.wavAdd --json before or after a command for one machine-readable result containing the schema,
job ID, status, typed artifacts, and provenance paths. jobs --status recoverable inspects the
same persistent job ledger used by the desktop recovery center. Generation commands remain
model-gated by default; --demo is required for the explicitly labeled MIDI and SFX fallback
paths. provenance reports whether an artifact is byte-replayable, semantically replayable, or
provenance-only, with the reason used by Project Manager before enabling re-render.
The reproducible engine evaluation harness uses fixed prompts, seeds, durations, and languages. It records runtime/model provenance, latency, RAM/VRAM, failures, audio loudness/true peak, artifact hashes, and a separate blinded listener rubric. Run the report-only fixture pass with:
py -3.12 tools/evaluate_engines.py --output evaluation-report.json --artifact-dir evaluation-artifactsPass a project runner using --runner module:function for real engine measurements;
the default runner deliberately marks cases as skipped and never gates a release on FAD.
Optional AI runtimes use platform-specific, SHA-256-locked profiles for CPython
3.12. Prepare a wheelhouse while connected, then the installation command
operates with --no-index and writes a complete CycloneDX 1.7 SBOM:
py -3.12 tools/dependency_profiles.py download windows-cpu --wheelhouse C:\SlunderWheelhouse\windows-cpu
py -3.12 tools/dependency_profiles.py install windows-cpu --wheelhouse C:\SlunderWheelhouse\windows-cpu
py -3.12 tools/dependency_profiles.py smoke windows-cpuAvailable locks cover Windows/Linux CPU and CUDA 12.6 plus Apple Silicon MPS. Torch-DirectML is intentionally disabled: Microsoft documents support only through PyTorch 2.4.1, below the PyTorch 2.10.0 security floor.
| Module | Description | AI Engine |
|---|---|---|
| Song Forge | ACE-Step 1.5 XL Turbo song generation, reference covers, source repaint/extend, stitched long-form songs, persistent retry/resume queue, and recovered vocal-stem export | ACE-Step, Demucs |
| Lyrics Engine | AI-powered lyrics writing with 33 genre templates, advisory rhyme/cadence feedback, and sortable history | Llama 3.1 8B |
| MIDI Studio | Piano roll editor with quantize/swing/humanize tools, CC lanes, text-to-MIDI composition, groove-template drums, and chord chart export | MIDI-LLM |
| Vocal Suite | Singing synthesis, humming-to-MIDI lyric melody generation with verified Enhanced LRC export, consented voice-profile onboarding, selectable Demucs/MDX/MDXC/Roformer stem separation, target-DAW stem naming/export, and vocal auto-tune pitch correction | DiffSinger, Audio Separator |
| Stem Separation | Isolate vocals, drums, bass, and other instruments | Demucs (htdemucs) |
| SFX Generator | Text-to-sound-effect generation | Stable Audio Open |
| Mixer | Project-rate mixing with validated resampling, channel normalization, dynamic EQ, Mid/Side trims, reference matching, mastering, and validated DAWproject export | Built-in DSP |
| AI Producer | Cancellable staged production with verified output, explicit demo/degraded states, and retry | Orchestrator |
| Model Hub | Download, verify, activate, deactivate, switch, filter, sort, and identify built-in core AI models | HuggingFace Hub |
| Projects | Searchable name/notes, date/name sorting, collision-safe asset imports, recoverable library index, version history, provenance tracking, and validated DAWproject export | — |
┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ AI Producer │───>│ Lyrics Engine│───>│ Song Forge │───>│ MIDI Studio │
│ (One Prompt)│ │ (33 genres) │ │ (ACE-Step) │ │ (Piano Roll)│
└──────────────┘ └──────────────┘ └──────────────┘ └──────┬───────┘
│
┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────▼───────┐
│ Export │<───│ Mixer │<───│ SFX Generator│ │ Vocal Suite │
│ (WAV/FLAC) │ │ (Mastering) │ │(Stable Audio)│ │(DiffSinger) │
└──────────────┘ └──────────────┘ └──────────────┘ └──────────────┘
Every module can route audio to any other module. Generate a song in Song Forge, separate stems in Vocal Suite, add SFX, mix everything in the Mixer, and export a mastered track. Engine actions expose their required model/profile and declared output types before they run; Model Hub keeps installed, verified, active, and failed states separate, while algorithmic MIDI/SFX fallbacks remain explicitly labeled demo outputs. RVC conversion and GPT-SoVITS cloning currently fail closed until verified local inference adapters are bundled; they never synthesize placeholder audio. Generated and exported assets write adjacent .provenance.json sidecars with app version, prompt/lyrics, seed, model revision/hash metadata, source paths, and render parameters so projects can be audited or reproduced. Project imports store every asset and sidecar under its stable asset ID, preserving same-named originals; Project Manager can rebuild a damaged or incomplete index from valid project folders and timestamped metadata backups. Long-running generation and model-download jobs persist queued/running/completed/failed/cancelled/recoverable state so interrupted sessions can show what needs recovery on restart. Settings and project files use versioned schemas with timestamped backups before migrations, repairs, and saves. Primary creative workflows expose screen-reader names, descriptions, high-contrast focus rings, and predictable tab traversal. Accessibility is gated by tests rather than asserted: every palette text/surface pair must clear WCAG 4.5:1, the focus ring must clear 3:1 on every surface and no stylesheet rule may hide it without a replacement, the waveform is fully keyboard-operable (arrows seek, Page Up/Down scrub, Home/End jump, M switches view) and announces its position as a value, notifications are mirrored into a persistent status line so timed toasts have a non-timed equivalent, and no view may demand more width than a 1024x768 display. For songs over 2 minutes, Song Forge can render structured sections separately and stitch them with crossfades for more stable long-form arrangements; completed Song Forge renders also attempt a Demucs vocal-stem recovery and expose a separate vocals-only route when recovery succeeds. Seed Explorer renders nearby seed/timestep-shift variations from the current lyrics and style prompt so you can compare takes before committing to a full arrangement. Reference Track analysis maps an audio fingerprint to ACE-Step tags for one-click style conditioning, Genre Fusion blends two template tag sets into weighted hybrid prompts, and Voice Cloning validates 10-30s GPT-SoVITS reference samples before saving reusable voice profiles with owner, consent source, language, permitted-use, and sidecar provenance metadata.
The open project autosaves on the interval configured in Settings > General, but only when it actually has unsaved changes. Each autosave writes a versioned snapshot alongside manual ones, and Project Manager can preview any stored version's name, tempo, key, assets, and mixer track count before restoring it. Restoring first snapshots the current state as a pre-restore version, so a restore can be undone. Retention is bounded by Settings > Kept Project Versions: the oldest autosaves are pruned first, manual versions next, and pre-restore versions are never pruned.
Settings > Recovery Center lists every recovery artifact - job records, per-job logs, crash logs, settings backups, and project versions - with its size and retention policy. Preview Cleanup always runs as a dry run and removes nothing; Clean Now applies the age, count, and size limits. Active jobs, recoverable jobs, the current crash log, the newest project version, and pre-restore snapshots are never removed.
Secrets such as the HuggingFace access token are stored in the operating-system credential service — Windows Credential Manager, macOS Keychain, or the Linux Secret Service — and never in the config JSON, its timestamped backups, logs, or diagnostics. Settings names the backend in use, and states plainly when no credential service is available instead of falling back to plaintext. A token found in an older plaintext config is moved into the credential service and removed from the config and its existing backups once the store confirms it can be read back.
Settings can export a redacted health report ZIP with app/dependency versions, GPU and ffmpeg status, model cache state, settings repair status, crash log metadata, and recent failed jobs. HuggingFace tokens are always redacted, and job prompts/lyrics stay out of the report unless the private-input opt-in is enabled.
The app chrome, Settings, Lyrics, and Vocal Suite controls use a catalog-backed locale layer. Settings > Appearance > Interface Language supports English, Arabic with right-to-left layout, and a pseudo-locale for layout QA; the selected locale is restored on restart. Settings > Appearance > Default Lyrics Language feeds Quick lyrics prompts, Guided lyrics metadata, and new GPT-SoVITS voice profile language defaults where supported. MIDI Studio supports explicit chord-progression priors such as I-V-vi-IV and ii-V-I for text-to-MIDI prompts and fallback generation. MIDI Studio also includes selectable drum groove templates with swing timing, snare ghost notes, and velocity humanization for generated GM drum tracks, .chordpro and .crd chord chart export with optional pasted lyrics, and piano roll editing tools for quantize, swing, velocity humanize, and MIDI CC automation lanes. Vocal Suite includes a Lyric Melody tab that converts hummed audio into provenance-tracked MIDI, aligns pasted lyrics to detected notes, and can render a routed DiffSinger vocal when a model is loaded. The Auto-Tune tab writes routed, provenance-tracked WAV files with adjustable pitch correction toward the nearest semitone.
Mixer converts every imported track to the configured project sample rate with deterministic polyphase resampling and explicit stereo normalization before calculating duration or summing. All lossless engine, mixer, and trim artifacts use the shared settings-aware writer, which preserves the requested sample rate, channel layout, and 16/24/32-bit WAV capability. The mastering true-peak meter intentionally oversamples only for measurement; it does not introduce a second processing resampler. Mixer can analyze each stem, infer a stem role from the track name, and apply local dynamic EQ suggestions with per-band gain, frequency, Q, and reasoning before mastering/export. It can also trim Mid and Side gain independently before limiting, load a reference track, match the final master to its integrated LUFS, report EBU Mode momentary (400 ms), short-term (3 s), and integrated readouts, and target streaming, podcast, EBU R128, broadcast, cinema, or loud CD delivery levels.
Project Manager exports the open project's existing audio assets as a .dawproject archive, while Mixer exports the current track buffers with their volume, pan, mute, and solo state. Both paths run off the UI thread, use collision-safe media names, and reopen the archive through Slunder's structural validator before reporting success. Archive validation and extraction reject traversal, absolute-path, duplicate, symlink, and oversized entries; recoverable trash validates manifest paths before restoring anything. Project and job schemas retain unknown fields so newer files can round-trip through an older shell, while malformed roots are quarantined or reported through repair diagnostics. Vocal Suite remix export preserves the separated stems' native sample rate and records the source paths and rate in its provenance sidecar.
Vocal Suite's Stem Mixer can export each separated stem as a native-rate WAV with a persisted target-DAW naming template (Generic, Ableton Live, Bitwig Studio, Cubase, Logic Pro, Pro Tools, Studio One, or FL Studio). Each delivery file gets its own provenance sidecar and stable track order; the selected convention is also available in Settings > Output. Live conversion and cloning progress labels describe the actual operation and do not call an activated-model run a demo.
Optional OSC transport control is available in Settings > OSC Control. It is disabled and
loopback-only by default, accepts direct messages in the versioned /slunder/v1 namespace, and
requires both an explicit LAN opt-in and matching IPv4 host/CIDR allowlist entries for network
sources. Packet-size and per-source message-rate limits apply before transport dispatch.
Optional MIDI controller mapping is available in Settings > MIDI Controller. It is disabled by
default and maps common CC/note controls to transport, the selected Mixer track's volume/pan/
mute/solo, and Piano Roll quantize/swing/humanize actions. Install mido plus a backend such as
python-rtmidi to expose local input ports; without that optional backend, mappings remain safely
stored and the rest of the app remains usable.
Generated and exported provenance sidecars also declare whether their operation is byte-replayable,
semantically replayable, or provenance-only; Project Manager and the headless provenance command
use that capability and its reason before offering a re-render.
| Preset | Target LUFS | Character |
|---|---|---|
| Balanced | -14.0 | Neutral, general purpose |
| Loud / Radio | -11.0 | Compressed, bright, competitive loudness |
| Warm / Analog | -14.0 | Enhanced lows, rolled-off highs, narrow stereo |
| Bright / Crisp | -14.0 | Enhanced highs, mid presence, wide stereo |
| Hip-Hop / Trap | -12.0 | Heavy sub-bass, punchy compression |
| Cinematic | -16.0 | Dynamic range, wide stereo, gentle compression |
| Lo-Fi | -16.0 | Rolled-off highs, heavy compression, narrow |
| Streaming (Spotify) | -14.0 | Optimized for streaming platform normalization |
Mixer exposes Mid and Side trim controls from -6 dB to +6 dB. The mastering chain applies those trims after stereo-width processing and before limiting/LUFS normalization.
| Target | LUFS | Use |
|---|---|---|
| Streaming | -14.0 | General music streaming normalization |
| YouTube | -13.0 | Online video music delivery |
| Apple Music | -16.0 | Apple Sound Check-style music delivery |
| Podcast stereo | -16.0 | Spoken-word stereo podcast delivery |
| EBU R128 | -23.0 LUFS / -1 dBTP | Broadcast programme loudness and true-peak ceiling |
| Broadcast | -24.0 | ATSC A/85-style broadcast delivery |
| Cinema dialog | -27.0 | Dialogue-oriented cinema delivery |
| CD / loud master | -9.0 | Loud physical or club master delivery |
EBU R128 is the shared target from EBU R 128 v5.0; its true-peak measurement follows ITU-R BS.1770-5 Annex 2.
Models are downloaded on-demand through the Model Hub. Nothing downloads until you need it.
| Model | Size | Module | Required |
|---|---|---|---|
| ACE-Step 1.5 XL Turbo (DiT) | ~10.4 GB | Song Forge | Recommended |
| Llama 3.1 8B Q4_K_M / Q8_0 | ~4.9 / 8.0 GB | Lyrics Engine | Recommended / high quality |
| DiffSinger (ONNX) | ~500 MB | Vocal Suite | Optional |
| RVC v2 | ~200 MB/voice | Voice profile metadata only; verified adapter required | Unavailable |
| Demucs (htdemucs) | ~300 MB | Stem Separation | Optional |
| Stable Audio Open | ~3 GB | SFX Generator | Optional |
All models run entirely on your local machine. No cloud APIs, no subscriptions, no data leaves your computer.
When Offline Mode is enabled in Settings, all HuggingFace API calls and model downloads are blocked — models load from local cache only or fail with an explicit offline error. C2PA signing is local-only when its optional RFC 3161 timestamp URL is empty; entering a timestamp URL is an explicit network opt-in, is rejected by Offline Mode, and records timestamp_mode: rfc3161_network in the signed export's provenance sidecar.
Model Hub cards show each model's license, gated/token status, and commercial-use status. Generated and exported provenance sidecars carry source model license policy forward, and Song Forge export warns when a source model is limited, non-commercial, or governed by model-specific terms.
Built-in Hugging Face downloads use reviewed immutable commit revisions and hash every cached file before loading. Model Hub can check upstream sources on demand, display bounded release notes for an immutable target SHA, validate the complete target cache before activation, and retain the prior verified cache for one-click rollback. When a publisher includes an OpenSSF Model Signing (OMS) detached signature, the locked model-signing verifier checks it against the local model file manifest before any loader runs; Model Hub and provenance show the signer result, while a model without a published signature is explicitly marked unsigned. Remote model code is disabled by default; models that declare custom code or pickle-backed weights require an explicit Model Hub warning and consent scoped to that exact revision.
Model Hub also records task labels, a dated measurement basis, and a VRAM tier for every built-in model. Its recommendations and hardware filter use the detected accelerator and the same registry VRAM estimate used by activation, calling out CPU fallback instead of presenting an oversized model as a GPU fit.
Lyrics GGUF models expose paired Q4_K_M and Q8_0 variants where the upstream snapshot provides both files. Q4 uses less disk and VRAM; Q8 preserves more quality at a larger footprint. The optional core.model_variants benchmark contract runs fixed prompts and records a caller-supplied quality score, token throughput, exact file size, RAM, and peak VRAM so local measurements stay separate from upstream catalog estimates.
After downloading a variant, run py -3.12 tools/benchmark_model_variants.py --model-id llama-3.1-8b-q8 to emit its measurement as one JSON record; add --output path.json to persist it.
Model downloads and model-bearing inference workers use central, cancellation-aware admission limits (two downloads and one GPU-bound inference by default), so parallel requests queue safely while existing resumable downloads and SHA-256 verification remain unchanged.
Song-generation backends are declarative registry entries: each active entry pins its source revision, portable local-mirror identity, loader, and adapter operations. Non-commercial or unlicensed generator weights fail closed until the exact source/revision terms are accepted. The pinned HeartMuLa OSS configuration is staged behind a separate runtime profile and remains disabled until that compatible environment is available.
| Component | Minimum | Recommended |
|---|---|---|
| OS | Windows 10 / Linux / macOS | Windows 11 / Ubuntu 22.04+ |
| Python | 3.11 | 3.12 |
| RAM | 8 GB | 16 GB+ |
| GPU | None (CPU mode) | NVIDIA 16GB+ VRAM for the ACE-Step XL path; 4–8GB tiers fit smaller models |
| Disk | 2 GB (app only) | 45 GB+ (full model registry) |
GPU acceleration requires PyTorch with CUDA support. The app runs on CPU without any GPU, but generation will be slower. The Model Hub explains the published VRAM tier for each task and filters recommendations against the detected device; an 8GB CUDA card can run lower-VRAM models, while ACE-Step XL is presented as a 16GB CPU-offload tier.
Settings are stored in ~/.config/SlunderStudio/ (Linux/macOS) or %APPDATA%/SlunderStudio/ (Windows).
The configured output directory is used by generated artifacts and renders; the default export
format, sample rate, bit depth, GPU index, MIDI tempo, Song Forge defaults, and mastering stage
switches are consumed by their respective runtime paths. The model-cache limit is an admission
cap: downloads that would exceed it are refused with an actionable message. First-run onboarding
can save the output folder and experience level, securely collect a HuggingFace token before a
gated download, and carry a selected core model into Model Hub for review or download. It can be
reopened from Settings after the first run; choosing Skip leaves setup incomplete.
Content Credentials are optional and off by default. Settings > Output can enable C2PA manifests for WAV, FLAC, and MP3 exports after you select your own PEM claim-signing certificate chain and unencrypted P-256 private key. Slunder Studio reads the key only for that export and never generates, copies, or stores the key contents. OGG and Opus exports remain unsigned because the current bundled C2PA binding does not embed those formats. Signed exports carry the creation action, a C2PA AI-disclosure assertion when a model was used, and a digest binding the manifest back to the adjacent provenance sidecar. Settings > Output also accepts an optional RFC 3161 timestamp URL: leave it empty for local-only signing, or enter it when an online timestamp is intended. Offline Mode rejects a configured timestamp URL before signing, and timestamp authority failures remove the incomplete artifact instead of presenting it as a signed export.
SlunderStudio/
├── settings.json # App preferences
├── voice_bank.json # Voice model profiles
├── projects/ # Saved projects with version history
├── models/ # Downloaded AI models
├── voices/ # Voice model files
└── generations/ # All generated outputs
├── songs/ # Song Forge output
├── lyrics/ # Lyrics Engine output
├── midi_studio/ # MIDI generation output
├── midi_renders/ # FluidSynth renders
├── vocals/ # DiffSinger output
├── voice_convert/ # RVC output
├── voice_clone/ # GPT-SoVITS output
├── stems/ # Demucs separation output
├── sfx/ # SFX Generator output
└── ai_producer/ # AI Producer pipeline output
Create a standalone executable from a temporary, hash-locked build environment:
py -3.12 build/build.py --clean-env
py -3.12 build/build.py --clean-env --onefile # Single .exe (Windows)The --clean-env build installs requirements-lock.txt and the pinned
build/requirements-build-lock.txt with --require-hashes, rejects extra
installed distributions, excludes test/cloud/unused Qt modules, and audits the
one-folder bundle before it can be packaged. It fixes PYTHONHASHSEED and
SOURCE_DATE_EPOCH, disables UPX, normalizes archive timestamps, and writes a
CycloneDX 1.7 inventory of the files that actually ship as
SlunderStudio.cdx.json inside the bundle. The Windows smoke check always uses
the private non-input virtual-display isolation helper. The build removes stale
dist/ outputs, resolves the platform icon, verifies one onedir process or the
expected onefile bootloader-parent/child tree, and writes dist/SHA256SUMS.txt.
The default build also creates
dist/SlunderStudio-vX.Y.Z-<platform>-<arch>.zip beside dist/SlunderStudio/.
Every version string — the window title, settings and project schemas,
provenance sidecars, the README badge, artifact names, and the embedded Windows
file version — comes from core/version.py.
Releases are unsigned. There is no code-signing step and none will be added. Windows SmartScreen will warn on first run of an unsigned executable; choose "More info" then "Run anyway", or verify the download against SHA256SUMS.txt first.
Windows releases are published on the GitHub Releases page.
Each release includes an unsigned ZIP and SHA256SUMS.txt; verify the checksum before running the app.
SlunderStudio/
├── main.py # Entry point with dependency diagnostics
├── core/ # Core infrastructure
│ ├── audio_engine.py # Playback engine (sounddevice)
│ ├── audio_export.py # WAV/FLAC/MP3 export
│ ├── chord_chart.py # MIDI chord inference and ChordPro/CRD export
│ ├── evaluation.py # Fixed-case engine measurements and listener rubric
│ ├── lyrics_db.py # Lyrics database with search
│ ├── mastering.py # DSP mastering chain
│ ├── midi_utils.py # MIDI I/O (pretty_midi wrapper)
│ ├── model_manager.py # HuggingFace model downloads
│ ├── disclosure.py # AI disclosure and human-authorship reports
│ ├── device.py # Configured CUDA/MPS/CPU device selection
│ ├── provenance.py # Generation sidecars and project metadata
│ ├── project.py # Project save/load/versioning
│ ├── settings.py # Persistent settings
│ ├── voice_bank.py # Voice profile management
│ └── workers.py # Background inference workers
├── engines/ # AI engine wrappers
│ ├── ace_step_engine.py # ACE-Step song generation
│ ├── ai_producer.py # One-prompt pipeline orchestrator
│ ├── audio_analyzer.py # BPM/key/loudness analysis
│ ├── demucs_engine.py # Stem separation
│ ├── diffsinger_engine.py # Singing voice synthesis
│ ├── fluidsynth_engine.py # MIDI-to-audio rendering
│ ├── lyrics_engine.py # LLM lyrics generation
│ ├── lyrics_templates.py # 33 genre template definitions
│ ├── midi_llm_engine.py # Text-to-MIDI generation
│ ├── melody_extractor.py # Humming-to-MIDI lyric melody extraction
│ ├── rvc_engine.py # RVC + GPT-SoVITS voice engines
│ ├── vocal_tuning.py # Vocal auto-tune pitch correction
│ ├── sfx_engine.py # Stable Audio Open SFX
│ └── style_tags.py # ACE-Step style tag database
├── ui/ # PySide6 interface
│ ├── main_window.py # Main window with sidebar navigation
│ ├── theme.py # Catppuccin Mocha dark theme
│ ├── onboarding.py # First-run wizard
│ ├── song_forge_view.py # Song generation page
│ ├── lyrics_view.py # Lyrics writing page
│ ├── lyrics_editor.py # Rich lyrics editor
│ ├── midi_studio_view.py # MIDI composition page
│ ├── piano_roll.py # QGraphicsView piano roll
│ ├── midi_mixer.py # MIDI track mixer
│ ├── vocal_suite_view.py # Vocal synthesis page
│ ├── stem_mixer.py # Demucs stem mixer
│ ├── sfx_view.py # SFX generation page
│ ├── mixer_view.py # Multi-track mixer + mastering
│ ├── ai_producer_view.py # AI Producer page
│ ├── project_manager.py # Project browser
│ ├── model_hub.py # Model download manager
│ ├── settings_view.py # Settings page
│ ├── waveform_widget.py # Audio waveform display
│ ├── mood_curve_editor.py # Mood/energy curve editor
│ ├── reference_panel.py # Reference audio panel
│ ├── seed_explorer.py # Seed variation explorer
│ ├── batch_view.py # Batch generation
│ └── toast.py # Toast notifications
├── assets/templates/ # 33 genre JSON templates
├── build/build.py # locked PyInstaller packaging
├── build/requirements-build.txt
├── build/requirements-build-lock.txt
├── requirements.txt # Dependencies
└── LICENSE # MIT License
Q: Do I need a GPU? No. Everything runs on CPU. A CUDA-capable NVIDIA GPU accelerates supported models; 4–8GB cards fit lower-VRAM models, while the default ACE-Step XL generation path is recommended at 16GB+ with CPU offload. Model Hub shows the exact fit and fallback status for the detected device.
Q: How much disk space do models need? About 15.3 GB for the recommended ACE-Step and Llama checkpoints. The full built-in model registry is approximately 38.4 GB. Models download on-demand — nothing installs until you request it.
Q: Can I use my own voice models? The Voice Bank can store consented RVC and GPT-SoVITS profiles and auto-detects standard files, but conversion and cloning remain unavailable until verified local inference adapters are bundled. No placeholder audio is produced.
Q: Is any data sent to the cloud? No. All processing is local. The only network traffic is model downloads from HuggingFace, which you initiate manually.
MIT License. See LICENSE for details.
Built by SysAdminDoc with Slunder.