Narrative Threads · Shortform Clip Demo

Scroll an entire movie down to 5-to-20 clips — ranked by beat, scored by model.

Voyager is a working preview of the narrative-thread pipeline: from a long-form Paramount film, extract playable vertical clips at the moments that matter, score them across Freytag beats, and compare what seven different rankers choose. Every clip is real — 1080×1920 H.264, word-level captions, VionLabs pan-and-scan.

60
Films in catalog
3,200+
Rendered clips across 7 rankers
99.9%
Word-level caption coverage
P1 · P2 · LF
Three clip passes · filter by shape

Explore

seven entry points into the project
Viewer

Clip Feed →

TikTok-style scrollable feed. 1,150 v6 clips across all 60 films. Sort by Freytag intensity, filter by movie, deep-link to any clip, cycle caption styles with v.

v6 default 1,150 clips · 60 films
A/B

Ranker Compare →

Side-by-side grid of up to 7 rankers picking from the same film. Toggle 4-col quality view (v1 / g3 / v5 / v6) or 7-col full view (+ v2 / v2b / v2c).

7-way A/B All variants loaded
Structure

Narrative Trees →

D3 hierarchy viewer for Pass-1 beats and Pass-2 sub-beats. Eleven flagship films plus a parameterized viewer for the rest. Click a leaf to play its clip.

D3 · interactive 11 films rendered
Compare

Tree Diffs →

Side-by-side per-film tree comparisons across v1 / mm3 / mm4 / mm4c variants. Synchronized horizontal zoom, click-to-play any beat with a 480p proxy. 12 films rendered — includes Titanic + Wolf ElevenLabs caption variants (EL v1 / EL v2).

17 diffs 5 movies · 5 episodes · 4 EL variants
Narrated

Reels →

Context-provider narration — narrator LEADS UP to each clip, then the clip plays with its original dialogue untouched. Same movie, three narrative threads: the love story, the disaster, and the treasure hunter's reckoning. Each reel walks a different path through the mm3 Pass 2 tree.

newest 3 threads · Titanic
Pass 1

Multimodal (mm3) →

Chapter-Llama recipe: interleave ASR with visual keyframe captions at shot cuts, then LLM-chapter with coverage enforcement. Wins pairwise vs v1 at 2:1.

new mm3 $0.30 / title
Architecture

Ranker Variants →

What changes between v1, v2/b/c, v5, v6, g3, mm3: model, thinking budget, selection rule, elite clip counts. Which one wins and where.

7 variants prompts · models · scores

Current winner

pairwise LLM judge, n=50 across 5 genres

mm3safe beats the v1 pilot at 2:1.

Pass 1 multimodal (mm3) replaces ASR-only beat extraction with ASR + visual captions at shot cuts, enforces per-3-minute-span coverage, and labels spoiler tiers for safe / teaser / full selection. Across Titanic, Wolf of Wall Street, Breakfast Club, Kill Bill 2, and Amistad, Gemini 2.5 Pro judged mm3safe the strongest clip set in 17 of 50 pairwise comparisons; the v1 pilot won only 8.

Read the multimodal breakdown →

17 / 50
mm3safe wins
8 / 50
v1 pilot

Ranker variants

clip counts = rendered MP4s in this demo
Variant Model Thinking Films Clips Render size Notes
v1 Gemini 2.5 Flash (pilot) off 60 300 8.7 GB Baseline. Score-aware elite picker.
v5 Gemini 2.5 Pro MEDIUM 59 295 6.1 GB Pass 2 model swap.
v6 Gemini 3.1 Pro Preview HIGH 60 1,150 34.4 GB default 20 clips / film. Headline quality.
g3 Gemini 3 Flash Preview default 59 295 6.2 GB Fast / cheap baseline.
v2 / v2b / v2c Deterministic (no Pass 2 LLM) n/a 11 165 3.8 GB CS34 merge-trim, multi-window, classifier. Archived.
mm3 safe/teaser/full Pass 1 multimodal (Gemini 2.5 Pro) HIGH 5 75 ~1.5 GB winner 37 titles scored; 5 published.
longform Architecture C (5–10 min reels) n/a 25 124 25 GB CS37/CS38 long-form reels.

Pipeline

how a film becomes a feed
1 · Coords
VionLabs pan-and-scan
Per-frame x_center from Periscope API. Cached locally.
2 · Source
Pull HD master
GCS cbs_ent/…/mux.mp4. Cut once, keep.
3 · Cut
vionlabs_cut.py
ffmpeg + numpy. 1080×1920, ~4s / clip. Deterministic.
4 · Captions
faster-whisper
medium.en, word-level timestamps.
5 · Upload
GCS + CORS
Paramount domain gated. Streams straight to browser.

Full pipeline walkthrough →