Variants · models · selection
Seven rankers. One clip set per film per ranker. Who picks best?
Every variant reads the same Pass-1 beat tree, runs the same Pass-2 sub-beat extraction, and returns a ranked
list of cut-points. What differs is the model, the thinking budget, and the
elite-picker rule. Side-by-side in Compare; matched across
individual films in Tree Diffs.
The variant roster (by pass)
Pass 2 — sub-beat short 30-90s clips · LLM-subdivided beats · default in feed
v1 pilot baseline
Gemini 2.5 Flash · no thinking
The original Sept pilot. ASR-only. Score-aware elite picker: sort sub-beats by score.overall, take top 5.
v5 pass-2 swap
Gemini 2.5 Pro · thinking=MEDIUM
Keep v1's structure; upgrade Pass-2 sub-beat LLM to Pro. Tighter sub-beat boundaries, less Hallmark prose.
v6 default
Gemini 3.1 Pro Preview · thinking=HIGH
Headline quality. 20 clips per film via Freytag-bucket picker (balances rising-action, climax, falling-action). Default in
the feed.
g3 fast baseline
Gemini 3 Flash Preview · no thinking
Cheap ceiling test. How close does Flash 3 get to Pro 3.1 if you don't pay for thinking? Useful for cost-quality curve.
v2 / v2b / v2c archived
Deterministic Pass-2 · no LLM
CS34 line: merge-trim (v2), multi-window (v2b), classifier (v2c). Rule-based. Kept for compare-grid baseline; no ongoing work.
mm3 safe / teaser / full pairwise winner
Gemini 2.5 Pro + visual keyframes · thinking=HIGH
Pass-1-informed Pass-2 sub-beats: Pass 1 gets visual keyframes + coverage-enforcing prompt; Pass 2 then subdivides into short clips. Three spoiler modes.
→ deep dive
Pass 1 — scene-length 60-600s whole-beat cuts · direct from Pass-1 beats · no sub-beat subdivision
mm3d · safe / teaser / full new
mm3 Pass-1 beats, rendered directly · Cloud Run
Scene-length (60-600s) clips cut directly from the mm3 Pass-1 tree — one clip per beat, no subdivision. Different product shape from 30s sub-beats.
→ open feed
Long-form 5-10 min highlight reels · story-arc grouped
longform architecture C
Story-arc grouped · 5–10 min reels
CS37 / CS38 experiment: group adjacent beats into 5–10 min highlight reels instead of isolated 30-second clips. Different product.
What changes vs what stays fixed
| Variant |
Pass 1 (beats) |
Pass 2 (sub-beats) |
Selection |
Thinking |
$ / film |
Clips |
| v1 |
Gemini 2.5 Flash | Gemini 2.5 Flash |
Top-N by score.overall |
off |
$0.10 | 5 |
| v5 |
Gemini 2.5 Flash | Gemini 2.5 Pro |
Top-N by score.overall |
MEDIUM |
$0.45 | 5 |
| v6 |
Gemini 2.5 Flash | Gemini 3.1 Pro Preview |
Freytag bucket × intensity |
HIGH |
$1.20 | 20 |
| g3 |
Gemini 2.5 Flash | Gemini 3 Flash Preview |
Top-N by score.overall |
off |
$0.15 | 5 |
| v2 / v2b / v2c |
Gemini 2.5 Flash | Deterministic (no LLM) |
Merge-trim / window / classifier |
n/a |
$0.08 | 15 |
| mm3 safe |
+ visual keyframes, coverage rule | Gemini 2.5 Pro (ASR-only) |
spoiler_level ≤ low |
HIGH |
$0.30 | 5+ |
| mm3 teaser |
(same as mm3 safe) | Gemini 2.5 Pro (ASR-only) |
spoiler_level = moderate |
HIGH |
$0.30 | 5+ |
| mm3 full |
(same as mm3 safe) | Gemini 2.5 Pro (ASR-only) |
all levels |
HIGH |
$0.30 | 5+ |
Cost is approximate; thinking=HIGH on Pro 3.1 dominates. Clips/film for v6 = 20 because of Freytag-bucket picker, not because the prompt asks for more.
Picking a variant
Which ranker should I look at?
“I want to see the strongest clip set per film, period.”
→
v6. 1,150 clips across all 60 films, Freytag-balanced, HIGH thinking. It's the default in
Feed.
“I want to see what the multimodal bet looks like.”
→
mm3 safe. Pairwise winner vs v1. 5 films published with tree diffs.
→ details“I want the cost-cheap baseline for benchmarking.”
→ g3 or v1. Both use Flash, no thinking. g3 is on the newer model.
“I want to see how clip selection differs without the LLM.”
→ v2 / v2b / v2c. Rule-based Pass 2. Good for isolating what the Pass-2 LLM actually adds.
“I want 5–10 min reels, not 30-second clips.”
→ longform. 124 reels across 25 films. Different product surface.
“I want to watch them all side-by-side for one film.”
→
compare.html. 7-column grid; toggle
a for quality-only (4-col) vs all-7.
Where each variant lives
| Variant | Clips manifest | Rendered MP4 path | Narrative tree |
|---|
| v1 |
feed_data_v1.js |
clips/<slug>_vion_<NN>.mp4 |
trees/<slug>.json |
| v5 |
feed_data_v5.js |
clips/<slug>_vion_v5_<NN>.mp4 |
trees/<slug>_v5.json |
| v6 |
feed_data_v6.js |
clips/<slug>_vion_v6_<NN>.mp4 |
trees/<slug>_v6.json |
| g3 |
feed_data_g3.js |
clips/<slug>_vion_g3_<NN>.mp4 |
trees/<slug>_g3.json |
| mm3 safe/teaser/full |
feed_data_mm3<mode>.js |
clips/<slug>_vion_mm3<mode>_<NN>.mp4 |
trees/<slug>_mm3.json |
Open the 7-way compare grid →
·
mm3 deep dive →