Every genre has a signature technique: short drama runs on two-person dialogue, martial arts on weapon state
Our technique piece covered the library-wide leaderboard. This one turns it around: which techniques does each genre prefer?
The measure is relative concentration (lift) = the technique's share within that genre ÷ its share across the library. A lift above 1 means the genre uses it more than average.
Where the data comes from
A full cross-tab of 28 techniques against the 18 genres with an adequate sample. Genre and technique tags are automatic candidates plus human review, unreviewed items excluded. Lift is relative, and genres with small samples (e.g. transition editing at n=43) fluctuate widely — flagged below.
Finding 1: Short drama is the only genre with a complete technique set
| Genre | Signature techniques (lift) |
|---|---|
| Short-drama dialogue (n=224) | Two-person dialogue ×2.5 (12.5%) · Listener reaction ×2.4 (4.0%) · Lip-sync ×2.2 (54.5%) |
| Martial arts (n=71) | Weapon state ×3.6 (39.4%) · Left/right staging ×1.9 · Wide shot ×1.8 |
| Fashion & beauty (n=94) | Static camera ×2.7 (6.4%) · Identity reference ×2.4 (22.3%) · Orbit ×2.0 (24.5%) |
| City & architecture (n=93) | Continuous single shot ×2.1 · Orbit ×1.7 · Aerial ×1.6 (33.3%) |
| Animation & games (n=83) | Weapon state ×2.4 · On-screen text ×2.4 (53.0%) · Reference image ×1.6 |
| Action & physics (n=70) | Weapon state ×2.3 · Left/right staging ×2.2 · Action causality chain ×1.9 (31.4%) |
| Landscape & nature (n=94) | Tracking ×1.4 · Environment reaction ×1.3 (43.6%) · Transformation ×1.2 |
| Everyday life (n=54) | Emotional performance ×1.6 (38.9%) · Prop interaction ×1.6 · Continuous single shot ×1.5 |
Short drama is the only genre with a complete dialogue-specific pair — two-person dialogue plus listener reaction ("listener reaction" means cutting to whoever is being spoken to), with a lift of 2.4. Across the library that technique appears just 21 times, a good share of them in short drama.
This explains why short drama is hard. It does not need one difficult technique; it needs a whole dialogue-blocking capability — who speaks, who listens, when to cut away. And in the public corpus, almost only short drama itself supplies that capability.
Finding 2: Martial arts has the most concentrated signature
Weapon state reaches a lift of 3.6 in martial arts (39.4% of martial-arts cases carry the tag, against 11.0% library-wide) — the highest lift of any genre-technique pair outside the small-sample transition-editing case.
The number is explanatory: the biggest problem with weapons in AI video is state inconsistency — a sword vanishing from a hand, a blade changing length, clipping when two weapons cross. So martial-arts creators handle "current weapon state" as a genre-specific technical step.
Finding 3: In reverse — the library's most common technique is badly under-represented in two categories
Environment reaction is the library's top technique (432 cases, 33.9%). Yet in some genres it falls far below average:
| Genre | Environment reaction share |
|---|---|
| Product ad | 20.3% |
| Food & dining | 20.4% |
| Visual design | 23.7% |
| … | |
| Library | 33.9% |
A 1.7x gap.
The other three leaders have their own gaps:
| Technique | Library | Lowest three genres |
|---|---|---|
| Close-up | 31.6% | Landscape 17.0% · Visual design 17.1% · Everyday life 20.4% |
| Emotional performance | 24.7% | City & architecture 12.9% · Action & physics 15.7% · Animals 17.1% |
| Lip-sync | 24.6% | Everyday life 7.4% · Animation 9.6% · Visual design 11.8% |
What these four have in common is being people-centred — which is why they dip together in genres whose subject is not a person: product ads, food, cityscape, landscape.
So the technique piece's conclusion ("the top four are all about people") needs a qualifier: it is a consequence of people-centred genres being the majority, not a universal law. Anyone making product or food content should discount those four techniques.
Finding 4: One small sample to be wary of
Transition editing (n=43) produces absurd lifts:
| Technique | Lift | Share |
|---|---|---|
| Listener reaction | ×8.5 | 14.0% |
| Three or more people | ×8.3 | 20.9% |
| Environment foley | ×5.3 | 37.2% |
A lift above 8 looks striking, but the base is tiny — listener reaction appears 21 times in the whole library, so 6 cases in transition editing are enough to produce a 14% share.
We recommend ignoring lifts that high outright unless the sample reaches at least 100. That is the threshold we apply to our own analyses in this series.
What you can do with this
If you write prompts for a specific genre: find that genre's signature techniques rather than copying the library leaderboard. Environment reaction is the overall leader but only 20% of product ads — copying it will produce an ad that reads like a short drama.
If you build prompt templates or tools: pre-configuring technique sets by genre is a data-backed design. Signature techniques differ per genre by a wide margin (lifts of 2–3.6), enough to warrant a template parameter.
If you are auditing your own output: check whether you have borrowed someone else's signature technique. Bolting a long emotional reaction onto a product ad usually means you started from a short-drama template.
Method and limitations
- High lifts in small-sample genres are unusable: transition editing (43), animals (41), tool comparison (44) fluctuate heavily. Only read genres with n≥90: short drama 224, product ads 192, character performance 149, music & dance 121, landscape 94, fashion & beauty 94, city & architecture 93.
- Tag boundaries are loose: one shot can be both "environment reaction" and "emotional performance", and annotators will differ. No inter-annotator agreement testing was done.
- Lift is relative: it says a genre uses a technique more than average, not that the technique works better in that genre. Effect is entirely out of scope here.
- Genres are multi-label: a case can belong to several genres, so genre samples sum to more than the case count and the lift denominator is not a strictly exclusive share.
- All figures use the 1,274 index records.
Reproduce it
The library supports multi-select genre and technique filtering, so any lift can be re-checked: https://cases.aishifu.shop/