Simple and Text-to-Music
Describe what you want in your own words. Safe defaults are filled in and visible, never hidden, so you can see exactly what is about to run.
Create
Generation with the controls a producer actually needs.
Start from a sentence or from a full parameter set — both paths reach the same engines, the same quality and the same follow-up workflows. Nothing is reserved for a higher plan.
- Text-to-music
- Custom
- Remix and cover
- Repaint
- Extract
- Lego layer
- Complete
- Batch
- Recipes
- Voice note
Modes
Ten ways in, one shared result object.
Whichever mode you start from, you end up with the same versioned track object — with its parameters, lineage, rights record and cost attached. Nothing is a side branch that cannot come back.
Custom
Take direct control of the musical decisions before a single second is rendered.
- Style and caption
- Lyrics with structure tags
- BPM, key and time signature
- Duration, voice language and instrumental
Remix and Cover
Start from an existing version or an uploaded reference and steer how far the result may move away from it with an audio strength control.
Repaint
Regenerate a selected region instead of the whole track. Four repaint modes plus crossfade and strength keep the rest of the arrangement intact.
Extract and Lego Layer
Pull out track classes from a mix, or generate a single new layer against what already exists, so an arrangement grows one decision at a time.
Complete, Batch and Recipes
Finish a partial mix, run a batch or auto-generation group under a fixed budget, or start from a saved recipe or a voice note.
Controls
Expert parameters, not a simplified dead end.
There is no separate Pro mode. Advanced and Expert controls open inside the same panel you started in, and they reach the same engines and the same quality as every other path.
Sampling and guidance
The parameters that decide how the model actually renders, validated against what the selected engine really supports.
- Seed and inference steps
- Guidance scale and CFG interval
- Shift and negative prompt
- Sampler and scheduler validation
DCW and engine specifics
Engine-level controls are exposed rather than hidden: DCW mode, wavelet and scalers, retake variance, flow-edit for caption, lyrics and morph, and legacy CFG behaviour.
Capability-driven UI
The interface is generated from the engine capability contract. A control that the deployed engine cannot honour is disabled and labelled, never silently ignored.
Engines
ACE-Step 1.5 and HeartMuLa sit behind one capability layer. Engine, model and adapter are pinned per job, so a platform-side model change never rewrites your old results.
Reproduction snapshot
Every run stores a complete snapshot — inputs, engine, model, adapters, parameters, seed and provider — so any version can be re-run or handed to a collaborator exactly as it was.
Assistant, not autopilot
The creator copilot suggests style, lyric structure, length and metadata with its reasoning visible. It proposes; you decide, and the parameters stay editable.
Words
Lyrics are a document, not a text box.
Lyrics carry their own history, structure and timing, and they travel with the track into every export.
Write, structure, improve
Draft lyrics with section structure, generate or improve passages with the assistant, and keep every revision as its own version you can return to.
LRC timing
Generate, edit, shift and validate synchronised lyrics, then export LRC alongside the audio for players, karaoke video layers and distribution.
Metadata suggestions
Title, tags, genre and description suggestions are proposed from the actual generation context rather than guessed from a filename.
Runs
Jobs that survive the tab you closed.
Generation runs on Soneth-managed infrastructure through a durable queue. Progress belongs to the job, not to the browser window that started it.
Real phases, not a spinner
The run reports the phases it is actually in — queued, model ready, rendering, post-processing, saved — with a separate technical diagnosis panel when something goes wrong.
Reconnect and cancel
Close the laptop, continue on the phone. Reconnect picks up live progress; cancel, cancel-all, retry and reset are first-class commands with clear failure diagnostics.
Cost before you spend
A cost preview appears before the paid submit. Budget warnings and a hard stop are enforced platform-side, and history shows provider, runtime and final cost per run.
Results session
Audition, compare, save or discard, and branch a follow-up run — all without losing the parent version or the session you were in.
Cover art in parallel
Cover generation runs as an independent job. It can never block, slow or fail the audio it belongs to.
History you can search
Every run stays searchable by prompt, engine, model, provider, cost and outcome, so a good result from three weeks ago is still findable.
The short version
- Engines
- ACE-Step 1.5 and HeartMuLa behind one capability contract
- Provider access
- Platform-managed only — no user API keys, no connected provider accounts
- Reproducibility
- Full parameter, model and rights snapshot on every version
- Cost control
- Preview before submit, budget warning, enforced hard stop