video_editing_poc/docs/cinematic-highlight-poc-plan.md

Local Cinematic Highlight PoC — Plan & Milestones

Owner: (Jaden) · Started: 2026-07-21 · Machine: Intel Mac (x86_64, 16 CPU, 32 GB, no GPU)

Objective

Produce ONE genuinely cinematic highlight from a local source video, end-to-end through the existing pipeline, using only local models resident in the service runtime: Piper (voiceover) + MusicGen (music) + AudioLDM2 (SFX), with deliberate visual treatment and pacing, and explicit human approval before any render.

A valid MP4 or a passing test is NOT success. Success = representative output passes measured media QA and a human creative review.

Non-negotiables (still in force during the PoC)

No external AI services in the media path · no placeholder silence/tones/OS say as a finished asset · no unlicensed assets · no rendering without explicit approval for the specific project · no inference-time network (models load from local dirs, HF_HUB_OFFLINE=1). NOTE: the user authorized model DOWNLOADS on 2026-07-21 (one-time provisioning), reversing the earlier no-download stance; inference stays offline.

Proven capability baseline (2026-07-21)

Model	Runtime	Status	Evidence
Piper en_US-lessac-medium	piper-tts 1.5.0	✅ works	real speech, 22.05 kHz
MusicGen musicgen-small	transformers 4.44.2 + torch 2.2.2	✅ works	5 s / 44.7 s CPU, mean −18 dB, 32 kHz
AudioLDM2 cvssp/audioldm2	diffusers 0.30.3	✅ works	3 s / 22.3 s CPU, mean −21.8 dB, 16 kHz

Environment traps (see memory local-model-runtime-intel-mac): torch capped at 2.2.2 (Intel-Mac), numpy<2, transformers must be 4.x, audiocraft/xformers unusable here, HF_HUB_DISABLE_XET=1 for downloads.

Plan

Phase 1 — Generate all three assets THROUGH the pipeline

• 1.1 Rewrite tools/local_asset_worker.py: music→transformers MusicGen, sfx→diffusers AudioLDM2 (resample to 48 kHz mono), voiceover→Piper (unchanged). audiocraft path removed. CLI + exit codes preserved.
• 1.2 Materialize models into stable models/ dirs via save_pretrained (tools/provision_local_models.py): models/musicgen-small (2.2 G), models/audioldm2 (4.2 G); models/piper (60 M) already present.
• 1.3 License/provenance .license.txt sidecars for each model marker file (config.json / model_index.json / voice .onnx). BOTH audio models are NON-COMMERCIAL (MusicGen CC-BY-NC-4.0, AudioLDM2 CC-BY-NC-SA-4.0) — flagged for production review. Honors the AssetLicensePolicy regular-file gate; no Java change.
• 1.4 Smoke test: voiceover 3.67 s/−15.8 dB, music 4.94 s/−12.0 dB, sfx 3.00 s/−20.6 dB — all 48 kHz mono, real signal, exit 0.
• 1.5 mvn -o verify → 245 tests / 62 classes / 0 failures / 0 errors / 0 skips (unchanged from baseline).

Phase 1 COMPLETE (2026-07-21).

Phase 2 — One approved end-to-end highlight (STOP before render for explicit approval)

• 2.1 Opt-in localpoc Spring profile added (application-localpoc.yml): venv python, model paths, offline, no bootstrap auto-start, heuristic visual, isolated PoC I/O dirs, render OFF, approval REQUIRED.
• 2.2 DJI source processed under localpoc (app started, ~4 min 4K analysis): wrote category.json (GENERIC_VLOG, conf 0.25) + highlight-candidates.json (3 windows: 27-39 s, 36-48 s, 81-93 s). Actual footage = blue Porsche Taycan aerial orbit (EV car reveal).
• 2.3 Hand-authored director/edit-plan.json: 3 x 9 s highlights (opening_hook / rising_energy / hero_payoff), grounded in the visible car; passes ALL validator constraints (containment, speed 1.0, duration bounds, VO budget, category match). Manual + local-only (no external AI).
• 2.4 RENDERED (operator-approved 2026-07-21). render-enabled override + approved.flag → local models generated 3 VO (Piper) + 3 music (MusicGen) + SFX (AudioLDM2) → 3 highlights rendered → project final.mp4 assembled (27.3 s, 51.7 MB). Flow ~8.6 min. All real local-model assets, no placeholders.
• [~] 2.5 QA measured (honest): PASS — 1920x1080 / H.264 / 30 fps / AAC 48 kHz; duration 27.3 s; integrated loudness -15.8 LUFS (target -16 ±1); no black frames; no long silence; overlays render & are grounded ("First light.", "Silent power."); VO intelligible. FAIL/weak — true peak: highlight_001 = 0.0 dBFS (clips; target ≤ -1.5 dBTP); h2/h3 = -1.4 dBFS (loudnorm TP limiting not enforced tightly). Visual "treatment" is essentially PASSTHROUGH — no visible cinematic grade/reframe (renderer maps any treatment to a generic look); overlays are plain white captions. LRA 16.9 (high). Pipeline's own qa-report passed because its grade/clip checks are structural. Verdict: genuine end-to-end local PoC, NOT yet "genuinely cinematic." Visual grade + true-peak limiting are the top Phase 3 items.

Phase 3 — Iterate to quality

Ranked from the first render's evidence:

• P3.1 Filmic grade in HighlightFfmpegRenderer.cinematicVisualFilter: S-curve + teal-orange colorbalance + eq + unsharp + vignette (replaces the weak fixed eq). Only the highlight renderer changed; multi-clip FfmpegEditRenderer untouched.
• P3.1b Beat-specific grading (a0b6023): grade now varies by story beat (read from the EditPlan style key, no EditDecision/serialization change) — opening=cool/soft (linear_contrast), rising=balanced (medium_contrast), hero=rich/warm (strong_contrast). Verified visually (calm→dramatic arc) + unit test.
• P3.2 True-peak safety: added alimiter=level=disabled:limit=0.72 after loudnorm in the audio mix. VERIFIED by re-render: project true peak 0.0 dBFS (clipping) → -1.7 dBFS; per-highlight -1.7/-2.7/ -2.8 (all <= -1.5 dBTP gate); integrated loudness -16.3 LUFS. (First pass at limit=0.79 gave -1.3 on h1, 0.2 dB over — tightened to 0.72 for inter-sample margin.) Test added.
• P3.3 Overlay styling: refined 48px caption, soft drop shadow, thin subtle border, smooth alpha fade in/out (0.4s ramps). Verified rendering; committed 24bd1d7. (Placement/safe-area unchanged.)
• P3.4 Real CV visual analysis via YOLOv8 (resident model, NO bootstrap script). Deps installed into an isolated .venv-local-cv; uvicorn worker run directly against the existing yolov8n.pt offline (YOLO_OFFLINE=True, telemetry sync off). Fixed a latent bug: the Java CV client negotiated HTTP/2 (h2c) which the HTTP/1.1-only worker mishandled (422/empty body) — pinned HTTP/1.1. Result on the DJI source: category GENERIC_VLOG@0.25 -> CAR_VLOG@0.95 (real car detection), measured blur 0.191 / exposure 0.95 (OpenCV), method local_cv_worker_opencv_yolo. Provider stays OPT-IN via runtime override (committed localpoc profile keeps the safe heuristic default; worker started manually). CAVEATS: yolov8n.pt is AGPL-3.0 (production blocker on this license alone; see yolov8n.pt.license.txt); loopback HTTP is flagged non-compliant for a certified path; Haar face detector gives false positives.
• [~] P3.5 Music/SFX + ducking review (measurement-based). FINDINGS: sidechain ducking IS implemented ([music_raw][voice]sidechaincompress with the configured threshold/ratio/attack/release; FFmpeg auto-splits [voice] so it both keys the duck and stays in the mix) and it fires (music drops during voice). Source audio is near-silent (-54 dB RMS) -> negligible. Loudness/true-peak stay in spec. OPEN: voice-vs-music balance cannot be validated or tuned by measurement alone — MusicGen produces different audio each run, so cross-render A/B is confounded, and this needs LISTENING. A blind +3.5 dB voice boost was tried and reverted (could not verify it helped; measurement suggested it did not). Per the "do not tune audio blindly" discipline, mix calibration is deferred to P3.6 (human review).
• [~] P3.6 Acceptance gate set up in docs/cinematic-highlight-acceptance-review.md: FROZEN technical thresholds + creative rubric (0–4, avg≥3.5 / no dim<3). Gate A (technical) measured = PASS on the canonical render (1080p/H.264/AAC 48k, −15.9 LUFS, TP −2.4/−2.5/−2.7 dBTP, no black/silence, all assets present; one marginal note on aggregate duration +0.32 s). Gate B (creative) requires the human review — cannot be self-assessed; awaiting reviewer scores.

Milestone log

• 2026-07-21: Models provisioned & individually proven (Piper, MusicGen, AudioLDM2). Plan approved. Phase 1 started.
• 2026-07-21: Phase 1 complete. Worker rewritten (audiocraft→transformers MusicGen + diffusers AudioLDM2), models materialized to models/ with license sidecars, all 3 asset kinds generate 48 kHz mono real audio through the worker, mvn -o verify green (245/0/0/0). Paused for review before Phase 2.
• 2026-07-21: Phase 1 committed (a95d1fa). Phase 2 up to approval gate: localpoc profile added, DJI source processed → candidates + category, grounded 3-highlight director plan authored & pre-validated. STOPPED — awaiting explicit render approval for dji_21230510013241_0091_d.
• 2026-07-21: Phase 2 rendered (operator-approved) — first full local source→final highlight reel.
• 2026-07-22: Phase 3.1 + 3.2 done & verified. Filmic grade + true-peak limiter. Re-render measured TP -1.7/-2.7/-2.8 dBFS (was 0.0 clipping), I -16.3 LUFS; mvn verify 245/0/0/0. Committed 588c652 (profile), 0fcfca9 (render quality).
• 2026-07-22: Phase 3.3 done & verified. Premium overlay styling (soft shadow, alpha fade). Committed 24bd1d7. Combined re-render (grade+limiter+overlays) verified: TP -2.5/-2.6/-2.8 dBFS, I -16.3 LUFS, 1080p/27.3s; mvn verify 246/0/0/0.
• 2026-07-22: Beat-specific grading done & verified (a0b6023). Grade varies per story beat; final re-render TP -2.3/-2.5/-2.8 dBFS, I -16.3 LUFS, 1080p/27.3s; mvn verify 247/0/0/0. Phase 3 visual/audio craft items complete.
• 2026-07-22: P3.4 real CV (YOLOv8) done & verified. Offline worker vs existing yolov8n.pt (no bootstrap). Fixed latent HTTP/2 client bug. DJI source now classifies CAR_VLOG@0.95 with measured blur/exposure and a real car label (was GENERIC_VLOG@0.25/filename heuristic). mvn verify 247/0/0/0. AGPL + loopback caveats recorded. Remaining: P3.5 music/SFX fit, P3.6 frozen thresholds + blinded review.

Phase 4 — Cinematic pass (after human review: "not genuinely cinematic")

Human Gate-B verdict on the Phase-3 output: functional, not cinematic. Direction chosen: dynamic editing, better music, NO voiceover, bolder finish. Done so far (commit 842f2a7):

• No voiceover — validator made VO optional; new plan carries none (music + SFX driven).
• 2.39:1 letterbox + film grain in the renderer; overlays raised above the bar.
• Slow-motion hero (plan: hero source span < target → 0.67x) + beat-specific bold grade.
• Stronger music direction — driving orchestral/hybrid prompts (still musicgen-small).
• Push-in / Ken Burns — prototyped via time-based crop zoom; INVALID (FFmpeg crop can't use t for w/h). Reverted. Redo with zoompan as a follow-up.
• Bigger music model — musicgen-medium download stalled + is very slow on CPU; deferred. Consider medium, one continuous score across the piece, or licensed production music.
• Re-run human Gate B on the cinematic cut. Verified: mvn 247/0; final −16.5 LUFS, TP −2.7 dBTP, 1080p, no VO. Deliverable: output/localpoc/highlight-projects/dji_21230510013241_0091_d/final.mp4.

2026-07-23 — Second source (bowling strike) + orientation/audio/duration fixes

• New source bowling_strike.mp4 (portrait phone footage) rendered e2e via montage; confirmed the DJI Phase-5 finding that source footage is the ceiling.
• Service-wide fixes applied (not just PoC), all with tests, mvn -o verify 254/0:
  • Orientation-aware rendering. FfmpegClipInspector now reads display rotation (side-data + rotate tag) and stores effective width/height; HighlightFfmpegRenderer.outputGeometry() renders portrait sources to a portrait frame and skips the 2.39 letterbox (landscape unchanged). Fixes portrait footage being stretched into landscape.
  • Audio mix. Source audio ([0:a]) ducked to −20 dB (−24 with narration) so the generated score leads instead of being buried; montage music cue raised to −9 dB.
  • Highlight duration constraint removed. Validator no longer enforces min/max (only positive/finite + sane speed + candidate containment); config bounds relaxed to soft candidate hints (min 2 / max 3600). A highlight can be as long as the story needs.
• E2e render flow captured in memory highlight-e2e-render-runbook (token-saver).

Phase 5 — Next steps (roadmap as of 2026-07-22, after montage mode)

To reach genuinely cinematic output (Gate B):

• P5.1 Better source footage (biggest lever; source is the real ceiling — parked car / industrial lot).
• P5.2 Automate shot selection — Tier 1 (DONE 2026-07-23). HighlightMontageDirector composes director/montage.json from measured motion (YDIF) + audio-energy (RMS) curves: setup → continuous action/tension → slow-mo payoff on the audio climax → resolution button, camera-whip tail trimmed. Wired into the scheduler behind auto-director-enabled (localpoc on). Verified: auto cut ≈ hand cut on bowling.
• P5.3 Better music — musicgen-medium/large (slow) or licensed production track; one produced score.
• P5.4 Polish: proper zoompan in-shot push-in; beat-synced cut lengths; better/optional voice. To make it a real system:
• P5.5 Resident local director (DONE 2026-07-23). Two tiers, both offline + auto: Tier-1 HighlightMontageDirector (measured motion+audio → shot list) and Tier-2 HighlightVisionDirector + tools/vision_caption.py (moondream2 captions the payoff → semantic overlay "STRIKE" + scene music). Wired into the scheduler (auto-director-enabled, vision-director-enabled; localpoc on). Plans are no longer manual. See docs/cinematic-quality-rules.md R9. To reach production (deferred until quality passes):
• P5.6 Formal Gate B human review (rubric in cinematic-highlight-acceptance-review.md).
• P5.7 Licensing blocker: MusicGen CC-BY-NC, AudioLDM2 CC-BY-NC-SA, YOLOv8 AGPL are non-commercial/ copyleft -> production needs commercially-licensed models/assets.
• P5.8 The hardening backlog below (security, CI/CD, containers, PostgreSQL, no-egress cert, repro build).

Director review of the bowling portrait cut (2026-07-23) — general craft gaps

Frame-by-frame + audio review. Verdict: technically clean (correct portrait, -15.5 LUFS, TP -2.8, no defects, score leads) but NOT yet top-tier cinematic. Craft gaps, all fixable as GENERAL pipeline features:

• P5.9 Adaptive exposure (DONE 2026-07-23). probeSourceLuma measures source frames (signalstats YAVG); exposureNormalizationFilter maps the measured mean to a target band (source-driven gamma); the grade is now exposure-preserving (black-lift, gamma≥1, brightness≥0, soft vignette). Bowling final went 63→~100 mean luma with a measured gamma of 1.098. General rules captured in docs/cinematic-quality-rules.md.
• P5.10 Life in held frames (DONE 2026-07-23). probeSegmentMotion (YDIF) measures per-shot motion; pushInAmount maps low motion → stronger in-shot zoompan push-in, high motion → gentle (floor keeps a little life). Applied before setpts so slow-mo survives. Bowling: static opening got 0.117 push, active celebration 0.084. See docs/cinematic-quality-rules.md R5.
• P5.11 Speed-ramp + crossfades. The cut into the slow-mo hero is an abrupt speed change and beats are hard cuts. Ease into slow-mo and add short crossfades on beat boundaries (renderer transition upgrade).
• P5.12 Bolder, animated overlay synced to the music hit. Current overlay is small plain white text; make it larger/bolder with a scale/fade punch timed to the payoff (general overlay upgrade).
• P5.13 Music climax sync. Align the generated score's peak to the hero beat's timeline position (pass the payoff timestamp into music prompting/trimming). Ties to P5.3.
• P5.14 Show the actual action. The edit never shows the pins fall (source is on the person). General: shot-selection (P5.2) should locate the true impact via audio-transient + motion, and source coverage should include it.

Deferred (until output-quality gate passes)

Production hardening: Spring Security/OIDC, PostgreSQL/Testcontainers, containers/K8s, CI/CD, distributed ops, digest-bound authenticated approval. Recorded, not deleted.