Repository navigation
Reelsmith v0.3.0
Optional sampled instruments, offline
Reelsmith can now render music from a local SF2 SoundFont bank instead of using the same procedural synth voices for every video.
The SoundFont bank is a separate, optional download (about 148 MB for FluidR3 GM). It is not included in the release archives and Reelsmith does not download it automatically. The executables still work without it using the existing procedural, file or no-music modes.
Three sampled palettes are available:
acoustic: piano, nylon guitar, upright bass and brushes.cinematic: strings, French horn and timpani.funk: electric piano, clean guitar, electric bass and drums.
Use music.kind = "soundfont", a local soundfont path and soundfont_style = "acoustic" | "cinematic" | "funk". MCP create_project / write_brief expose the corresponding music_soundfont and music_soundfont_style fields. See the sampled-music setup.
The renderer is pure Rust, using MIT-licensed RustySynth 1.3.8. The tested bank is FluidR3 GM v3.1, distributed under MIT. Download FluidR3_GM.sf2 from that release, keep its license notice and supply its local path. Tested SF2 SHA256: 74594e8f4250680adf590507a306655a299935343583256f3b722c48a1bc1cb0. Other banks need their own license check and the required GM presets. Missing or invalid banks fail clearly rather than silently falling back to the old sound.
Brief-driven musical controls
Optional calm, uplifting and tense directions set default harmony and tempo. Explicit BPM, MIDI root and custom progression are supported with validation; a root plus explicit progression is rejected rather than silently ignored. Per-beat/end-card energy changes note density and velocity in the sampled arranger, or density/gain in the directed procedural bed. Briefs can inherit project defaults.
This is a deterministic arranger with sampled instrument palettes, not an AI music model or automatic visual/content analysis. The person or host writing the brief chooses its musical direction. No paid service or network request is needed for synthesis after the bank is available locally. Old procedural defaults, external audio files and no-music mode remain available.
Validation
- Feature PRs #55 and #57; version PR #56. The first procedural profiles changed composition but not the sound source; #57 adds the sampled source.
- Current-head Linux x86_64 and ARM64 CI builds and teardown skill tests passed before feature merge. Eight parser/validation library tests passed locally; source reviews found no remaining demonstrated blockers.
- The downloaded feature CI Linux x86_64 binaries passed MCP schema, inheritance, explicit root/tempo, canonical/alias scene-energy, invalid-input brief preservation, real 6-second preview rendering and QA, cache reuse and energy-only invalidation.
- Three integrated-engine 6-second 360p previews passed delivered-file QA: acoustic -14.03 LUFS / -1.88 dBTP; cinematic -13.96 / -3.77; funk -14.02 / -1.99. Zero detected black frames. Contact-sheet pixels were inspected.
- Missing, empty and malformed banks, a valid bank missing required GM presets, fractional MIDI pitches and out-of-range transpositions were rejected. Bank mtime changes regenerated audio; bank content changes invalidated the full-resolution preview proof.
- Published v0.3.0 Linux x86_64 archive downloaded and SHA256 matched GitHub asset digest (
8c72bfe7b65d438ef4e2603a472473942c52977c54d15e56adaa527e50265381). Its actual executables reported 0.3.0 and passed MCP fields/inheritance, invalid-input preservation, a real 6-second sampled preview with QA, unchanged-audio reuse and energy-only regeneration. Contact-sheet pixels inspected. All four release targets built and passed MCP tool-discovery smoke checks.
Limits
The bank is user-supplied and is not embedded. SF2 reads are capped at 256 MiB, not a peak-memory guarantee or a claim of malicious-file parser hardening. Sampled renders support up to 180 seconds. The sampled bed plays across the cut with the final fade; legacy procedural drop/outro/seed variation is not implemented for this sampler. Scene energy is not a hard mute and exact beat-to-cut alignment is not promised.
Audio cache bank identity uses byte length plus mtime. Deliberately changing a bank without changing either may evade that cache; the preview proof hashes bank content. Objective QA does not prove musical taste. No speedup, HD caption readability or manual Windows/macOS/ARM runtime validation is claimed. The Linux test host emits a nonfatal weak GLIBC_2.39 warning; passing these checks is not a general compatibility guarantee.