Skip to content
Motif Motif

Roadmap

Where Motif Is Today

Section titled “Where Motif Is Today”

Motif has grown from a schema-only data format into an adaptive soundtrack workstation that plays real audio end to end — authored MIDI synthesis and generated beds, layered, transitioning, and audible in the Studio:

  • Composition: Clips with notes, instruments, scales, chords, motif transforms (30+ transforms in @motif-studio/clip-engine), intensity variants, and cue structures with timelines and section roles
  • Arrangement: Scenes with stem layers, gain control, mute/solo, section roles, intensity settings, and clip layer references
  • Adaptive Logic: Trigger bindings with condition evaluation (eq, neq, gt, gte, lt, lte, includes), deterministic priority-based scene resolution, and five transition modes — honoured both between scenes and inside cue playback, and fully cancellable (Stop always stops, even mid-transition)
  • Real-Time Playback: The playback engine is wired to the Studio — scene stems, synth clip layers, generated beds, stingers, crossfades, bar-sync and cooldown transitions all sound from the transport, with a mixer (buses, sends, per-stem FX inserts) underneath
  • Generated Cues & Library Packs: A catalog-driven pipeline turns AI-generated audio into playable packs — one scene, cue family (with a generation lock), and binding per cue, A/B takes with loudness normalization under a +6 dB boost cap, and full generation provenance on every record. See Generated Cues and Library Packs
  • Custom Samples: Import with filename-based source type inference, non-destructive trim, even and onset-based slicing, kit building, and pitched sample instruments
  • World Scoring: Motif families, score profiles with sonic identity parameters, cue families grouped by role, world map entries by context type, and derivation with seven transform types
  • Automation: Lanes with four interpolation curves, macros (intensity/tension/brightness/space), section envelopes with six shapes, and live performance capture with thinning
  • Recall and Reuse: Templates, snapshots, branches with lineage tracking, favorites, collections, and field-by-field entity compare
  • Studio: Full authoring UI with 18 screens (Next.js 15 + Zustand 5) on a single in-memory pack, including a runtime Preview whose state controls are derived from the loaded pack’s own bindings
  • Validation: Zod 4 schema validation, integrity auditing via @motif-studio/asset-index, and 1,709 tests spanning all 16 packages plus the Studio

Current Strengths

Section titled “Current Strengths”
  • Deep music model: The data model captures real musical intent — notes, scales, chords, motifs, intensity variants, cue timelines — not just audio file references
  • Deterministic runtime: Scene resolution is fully deterministic with priority-based binding evaluation and pack-order tie-breaking
  • Provenance-first generation: Every generated take records the exact prompt, seed, tempo, key, and job that produced it; prompt text is append-only, so a record never claims a prompt that didn’t produce it
  • Composition transforms: 30+ clip transforms enable rapid variant generation for adaptive layering
  • Library system: Snapshot, branch, compare, and template make creative iteration safe and traceable
  • Zero runtime dependency: The exported pack is plain JSON with no Motif runtime required

Known Gaps

Section titled “Known Gaps”
  • No take switching in the Studio UI — A/B takes both live in the pack and the runtime honours playbackDefault, but changing the default currently means re-running ingest
  • Ingest does not scale, and this is the biggest one — measured at ~1.2 takes/min (~43 s per take), so a 24-pack library is a ~6.7 hour build writing ~42 GB of 24-bit WAV. Three separate causes, each worth roughly an order of magnitude: there is no content-hash skip, so an unfiltered re-run rebuilds unchanged takes byte-for-byte; every take is transcoded to WAV when browsers decode FLAC natively; and the loop is fully serial with synchronous writes. Fixing any one makes library-scale practical. Until then, always pass --pack or --tier — see Generated Cues
  • No loop/seam construction yet — generated beds play as fixed-length loops without crafted seam points
  • No collaborative editing — single-user, single-file authoring only
  • No plugin/VST support — built-in instruments, sample playback, and generated beds only

Near-Term Phases

Section titled “Near-Term Phases”
  • Library breadth — ingest and audition the remaining catalog tiers; per-cue re-rolls where a take underperforms (extra seeds are cheap; quiet slow cues especially benefit)
  • Faster ingest — content-hash skip first (turns a re-run from hours into seconds), then serve FLAC instead of transcoding, then parallelise per-take work
  • Take curation UI — switch A/B/C takes per cue from the Studio instead of re-ingesting
  • Intensity-variant takes — calm/intense conditioned generations per cue family, complementing stem-mute intensity tiers
  • Runtime integration — packaging and game-state adapter guides for engines consuming exported packs, including generated beds

Long-Term Phases

Section titled “Long-Term Phases”
  • Deeper composition assistance — suggestion, not generation: AI proposes, the creator decides
  • Advanced sampler — granular synthesis, wavetable, and source design tools
  • Team and review workflow — multi-user commenting, approval gates, version management
  • Score QA — gameplay coverage analysis, dead scene detection, binding conflict reports
  • Platform and ecosystem expansion