Provisioned and tested the stronger Tier-2 judge on this Intel Mac (CPU-only):
- vision_caption_llamacpp.py: force CPU (-ngl 0 --no-mmproj-offload) because the
integrated GPU times out on the vision encoder (Metal command-buffer timeout);
extract the final assistant turn from the chat-templated output.
- docs/LOCAL-MODELS.md: the VERIFIED build+run recipe, incl. two real gotchas ->
Command Line Tools libc++ mismatch (add -isystem <SDK>/usr/include/c++/v1 to the
cmake flags, else ggml-base fails on <array>), and the Intel-GPU Metal timeout.
Verified result: where moondream described the bowling celebration as "standing in
a bowling alley", Qwen2.5-VL-3B says "raising their arms in a celebratory gesture"
(highlight-worthiness 1.0) and rates turn-around/anticipation low (0.2 / 0.15). End
to end, the director's judge now chooses time=10.5s score=1.0 (the celebration) vs
moondream's 13.0s score=0.15 (the turn-away). Model weights are gitignored under
models/ (Apache-2.0, provisioned offline).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The director's vision JUDGE is only as good as its model, and moondream2 can't
perceive some actions on hard footage (R15 ceiling). Make the captioner backend a
config choice so a stronger local VLM drops in with no code change:
- tools/vision_caption_llamacpp.py: same manifest->JSON contract as
tools/vision_caption.py, but backed by llama.cpp `llama-mtmd-cli` (GGUF). Runs on
this x86 CPU via AVX and bypasses the torch==2.2.2 / transformers 4.x trap
entirely (no PyTorch). Model/mmproj/binary paths come from env vars; fully
offline, serverless (per-frame CLI, mmap stays warm).
- editing.vision-caption-script selects the worker (default: moondream). The Java
HighlightVisionDirector now reads the configured script instead of a hardcoded
path -- nothing else changes.
- docs/LOCAL-MODELS.md: provisioning + enablement for Qwen2.5-VL-3B (Apache-2.0)
via llama.cpp; alternatives (Qwen3-VL, Gemma 3 4B). Honest note: likely improves
the bowling case but unverified until tested with real weights.
mvn verify: 293 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Choosing WHICH moment is the highlight is a question of meaning, not motion or
loudness, and there is no generic rule in measurement alone: on the real bowling
clip a camera turn-away has the highest motion AND is louder than the celebration.
Positional bands / thresholds only move which video breaks. So responsibilities
are now split, with zero per-video constants:
- HighlightMontageDirector.candidatePeaks: measurement PROPOSES the intensity
(motion+audio) local maxima, strongest-first, min-separated. No opinion on which
is the highlight; no band, no threshold.
- HighlightVisionDirector.rankDecisiveMoment: the vision model JUDGES each
candidate by highlight-worthiness (a celebration/goal outranks a loud turn-away
or an "about to..." build; anticipation is not the payoff). Highest score wins;
falls back to the strongest peak only if the model declines.
- composeMontageAt: builds the action segment (measured onset -> chosen peak ->
measured resolution that sweeps in the outcome+reaction) around the choice.
- MomentChooser interface makes the judge a drop-in: a stronger local VLM plugs in
with no director changes.
Removes the previous positional-band / semantic-weight heuristics.
Honest, verified ceiling (documented in R15): the judge is only as good as its
eyes. moondream2 perceives some actions (soccer: "kicking a soccer ball" -> the
goal is chosen correctly) but not others -- on distant portrait bowling footage it
describes every frame as "standing"/"walking" and never sees the celebration
(a posture prompt collapsed to a constant "Standing still"). When it can't
discriminate, candidates tie and it falls back to the loudest peak. This is a
model-capability limit, not a design flaw; the fix is a stronger VLM (drop-in).
mvn verify: 293 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- AGENTS.md: portable onboarding for any assistant (Codex/Claude/etc.) at the repo
root -- what the project is, read-order, build/test/run, hard constraints, honest
status, env gotchas. Makes the whole repo usable by any AI, not just Claude.
- docs/gate-b-review-bowling.md: the production-readiness scorecard. Gate A
(technical) measured = PASS; Gate B (human creative rubric) = PENDING. Output is
production-ready only when BOTH pass; this is the artifact a human fills to decide.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MPuJXQyAeWpFcTtcnxo1UN
Port the run-to-final flow and the local model runtime/version-trap facts from
Claude-specific auto-memory into repo Markdown so any assistant or human reading
the repo (Codex, Claude, etc.) has them. Runbook updated for the auto-director
(the plan is generated automatically now; manual authoring is an override).
Linked from the README.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MPuJXQyAeWpFcTtcnxo1UN
A slow-motion shot now decelerates smoothly instead of snapping to slow-mo:
speedRampSetpts builds a log-integrated setpts that ramps playback speed from
normal (1.0) down to below the target across the shot. The shot stays a SINGLE
segment (so the R5 per-shot push-in is preserved) and the existing -t pin keeps
the planned output duration. Normal-speed shots keep a plain constant setpts.
Unit-tested; verified in a real render (payoff carries the ramp, output valid).
Completes R6 (crossfades + speed-ramp).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MPuJXQyAeWpFcTtcnxo1UN
MusicGen's internal structure is uncontrolled, so the mix applies a deterministic
swell envelope to the score (volume='min(1,0.5+0.5*t/peak)':eval=frame): the music
amplitude rises from 0.5x to full over the run-up to the payoff, then holds. The
peak is the payoff (slow-motion) beat's timeline midpoint on the crossfade-
compressed timeline -- generic, driven only by the plan, no content assumptions.
No-ops when there is no slow-mo beat. Filter string unit-tested; verified in a real
render (swell peaks at 6.83s, output -16.1 LUFS).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MPuJXQyAeWpFcTtcnxo1UN
The vision director now captions several beat frames (two questions per frame in
one worker call: a discriminative description + a punchy label) and turns the
descriptions into a per-window "highlight-worthiness" curve via generic
emotion/action/idle keyword scoring (semanticScore/semanticCurve). The montage
director blends that curve with audio to place the payoff on the semantically
strongest moment; the payoff label becomes the bold overlay and the description
flavors the music. Everything is content-agnostic and fails soft to the measured
cut. Verified on bowling: the payoff moved onto moondream's detected celebration.
Honest limit: a small VLM on distant subjects is only weakly discriminative;
descriptive questions beat terse ones (which collapse to a constant answer).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MPuJXQyAeWpFcTtcnxo1UN
Overlays are now large (fontsize 84) with a thick outline + strong drop shadow so
they read on any background, and animate in: a snappy 0.18s alpha punch plus a
34px rise-up over 0.22s, with a soft ease-out. Placed on the payoff beat so the
entrance lands on the musical/edit accent. This presents the Tier-2 vision
director semantic caption ("STRIKE") boldly. Overlay style/animation asserted in
the existing overlay test.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MPuJXQyAeWpFcTtcnxo1UN
Beats now blend instead of hard-cutting (incl. the cut into the slow-mo payoff)
via an xfade + acrossfade chain (xfadeTimelineCommand). The timeline compresses
by (n-1)*xf, so shiftOverlayForCrossfade re-times overlays and the reported
duration is reduced to keep overlays, loudness mastering, and QA aligned. Opt-in
via editing.crossfade-seconds (0 = hard cuts default; localpoc 0.25), clamped to
half the shortest beat. Offset math + overlay shift unit-tested; verified on the
bowling cut (visible dissolve, STRIKE overlay stayed on the payoff).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MPuJXQyAeWpFcTtcnxo1UN
Single-pass loudnorm in the mix is only ~+/-2 LUFS accurate, so the auto-rendered
output could land quiet (e.g. -18.7 LUFS vs the -16 target). After the mix,
probeIntegratedLoudness measures the file, loudnessGainDb computes the corrective
gain, and masterLoudness applies it with a brickwall limiter for true peak. No-ops
when already on target or when the measurement is implausible; handles MusicGen
loudness variance. loudnessGainDb unit-tested.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MPuJXQyAeWpFcTtcnxo1UN
Rendering (whole service, driven by source measurements, not constants):
- Orientation-aware geometry: FfmpegClipInspector reads display rotation and
stores effective dims; HighlightFfmpegRenderer.outputGeometry renders portrait
sources portrait and skips the 2.39 letterbox on portrait (landscape unchanged).
- Dynamic exposure: probeSourceLuma measures the frames; exposureNormalizationFilter
maps the mean toward a target; the grade is now exposure-preserving (no crushed
subjects: bowling final went ~63 -> ~100 mean luma).
- Motion-adaptive in-shot push-in (zoompan), amount from per-shot YDIF.
- Audio mix ducks source audio under the generated score so it leads.
- Highlight duration is no longer capped (validator + config).
Automatic director (plans were hand-authored before):
- Tier 1 HighlightMontageDirector: composes the montage from measured motion (YDIF)
and audio-energy (RMS) curves -- setup, continuous action/tension, slow-mo payoff
on the audio climax, resolution button, camera-whip tail trimmed.
- Tier 2 HighlightVisionDirector + tools/vision_caption.py: a local, offline
vision-language model (moondream2) captions the payoff frame and augments the
montage with a semantic overlay ("STRIKE") and scene-informed music; fails soft.
- Wired into the scheduler behind auto-director-enabled / vision-director-enabled
(on in the localpoc profile).
Docs: cinematic-quality-rules.md (R1-R5, R9 both tiers), poc-plan milestones.
Tests: mvn -o verify -> 262 passing, 0 failures.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MPuJXQyAeWpFcTtcnxo1UN
The local-cv provider was never exercised end to end (its bootstrap was
prohibited), and hid a latent bug: the JDK HttpClient defaulted to HTTP/2 and
negotiated an h2c cleartext upgrade that the HTTP/1.1-only worker (uvicorn/h11)
mishandled by dropping the request body, so every call returned HTTP 422. Pin the
client to HTTP/1.1.
With this fix the resident YOLOv8 worker (run offline against the existing
yolov8n.pt, no bootstrap script) classifies the sample source as CAR_VLOG at 0.95
with measured OpenCV blur/exposure and a real car label, replacing the previous
filename-keyword GENERIC_VLOG fallback. Provider remains opt-in via runtime
override; the committed localpoc profile keeps the heuristic default.
Also ignore yolov*.pt.license.txt (provenance for the git-ignored weights).
mvn -o verify green (247 tests, 0 failures).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bg76sLc43Wc3j5ZcLkboYR
Record the operator-approved first render and the verified grade + true-peak
limiter results (TP -1.7/-2.7/-2.8 dBFS, I -16.3 LUFS) with commit references.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bg76sLc43Wc3j5ZcLkboYR
Phase 3 quality fixes for the highlight renderer (multi-clip FfmpegEditRenderer
left untouched):
- cinematicVisualFilter: replace the weak fixed eq with a deliberate filmic grade
(curves medium_contrast S-curve + teal-orange colorbalance + eq + unsharp +
vignette). Richer blue, warm highlights, tonal contrast. Still one uniform look;
beat/category-specific grading is future work (needs a validated grade enum).
- audio mix: add a brickwall limiter (alimiter limit=0.72) after loudnorm. The
first render clipped at 0.0 dBFS true peak; measured re-render now lands
-1.7/-2.7/-2.8 dBFS per highlight (all within the -1.5 dBTP gate), integrated
loudness -16.3 LUFS.
Test updated; mvn -o verify green (245 tests, 0 failures).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bg76sLc43Wc3j5ZcLkboYR
Replace the unusable audiocraft path (requires xformers, which has no Intel-Mac
build) with runtimes proven to work offline on this machine:
- music: transformers MusicGen (facebook/musicgen-small)
- sfx: diffusers AudioLDM2 (cvssp/audioldm2), resampled 16k -> 48k
- voiceover: Piper (unchanged), normalized to 48 kHz mono
The worker CLI contract and exit codes are preserved, so the Java
LocalAssetSynthesizer license gate and fail-closed behavior are unchanged.
Add tools/provision_local_models.py to materialize models into models/ from the
local HF cache with no network. Models and their license sidecars live under the
git-ignored models/ dir; both audio models are non-commercial (CC-BY-NC-4.0 /
CC-BY-NC-SA-4.0), recorded for later production review.
Add docs/cinematic-highlight-poc-plan.md tracking the PoC plan and milestones.
mvn -o verify: 245 tests, 0 failures/errors/skips (unchanged).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bg76sLc43Wc3j5ZcLkboYR
JSLMPR