AI video prompt corpus statistics
Direct answer: across 5,926 analysed AI video prompts from 1,196 authors, 85.8% specify camera language, 79.4% specify lighting, 98.3% carry a mood, and 52.9% state a duration. Tracking is the most common camera term (2,029 prompts) and 15 seconds is the dominant duration (71% of prompts that state one).
Computed 2026-09-11 from our own structured corpus (engine: deepseek-flash). Reproducible with pipeline/build_stats.py.
How many prompts specify each field?
| Field | Prompts | Share |
|---|---|---|
| camera | 5,082 | 85.8% |
| lighting | 4,705 | 79.4% |
| mood | 5,828 | 98.3% |
| duration | 3,136 | 52.9% |
| aspect ratio | 1,607 | 27.1% |
| negative prompt | 2,935 | 49.5% |
| consistency lock | 5,311 | 89.6% |
Camera and lighting are the most reliably specified fields: roughly 83% of prompts describe both. Aspect ratio is the least specified at 27.1% — creators rarely state it even though it changes the entire shot design.
Which camera movements are most common?
| Camera term | Prompts | Share of corpus |
|---|---|---|
| tracking | 2,029 | 34% |
| close-up | 1,951 | 33% |
| handheld | 1,370 | 23% |
| wide shot | 1,118 | 19% |
| slow motion | 836 | 14% |
| orbit | 739 | 12% |
| pull-out | 569 | 10% |
| aerial/drone | 568 | 10% |
| push-in | 551 | 9% |
| pan | 537 | 9% |
| zoom | 500 | 8% |
| low-angle | 492 | 8% |
| medium shot | 452 | 8% |
| shallow depth of field | 422 | 7% |
| POV | 402 | 7% |
| tilt | 379 | 6% |
| static/fixed | 375 | 6% |
| macro | 329 | 6% |
| slow push-in | 228 | 4% |
| one-take | 222 | 4% |
| dolly | 201 | 3% |
| whip pan | 162 | 3% |
| crane | 120 | 2% |
| establishing shot | 102 | 2% |
| overhead | 89 | 2% |
| speed ramp | 81 | 1% |
| match cut | 78 | 1% |
| rack focus | 67 | 1% |
| over-the-shoulder | 56 | 1% |
| high-angle | 52 | 1% |
| dutch angle | 17 | 0% |
Which lighting terms are most common?
| Lighting term | Prompts | Share of corpus |
|---|---|---|
| natural light | 1,397 | 24% |
| volumetric | 885 | 15% |
| warm tone | 879 | 15% |
| golden hour | 810 | 14% |
| dramatic lighting | 798 | 13% |
| soft light | 733 | 12% |
| neon | 639 | 11% |
| cinematic lighting | 577 | 10% |
| low-key | 543 | 9% |
| silhouette | 452 | 8% |
| practical lights | 415 | 7% |
| backlight | 367 | 6% |
| lens flare | 347 | 6% |
| cool tone | 319 | 5% |
| rim light | 310 | 5% |
| studio lighting | 167 | 3% |
| moonlight | 163 | 3% |
| hard light | 149 | 3% |
| practical lighting | 132 | 2% |
| spotlight | 86 | 1% |
| candlelight | 63 | 1% |
| high-key | 50 | 1% |
| theatrical lighting | 9 | 0% |
What durations do AI video prompts target?
Of the 3,136 prompts that state a duration, 71% specify 15 seconds. That single value dominates the corpus far more than any other field, which suggests 15s is the de facto standard clip length for this generation of models.
| Duration | Prompts | Share of stated |
|---|---|---|
| 15s | 2,224 | 71% |
| 10s | 278 | 9% |
| 30s | 123 | 4% |
| 12s | 110 | 4% |
| 14s | 55 | 2% |
| 8s | 45 | 1% |
| 16s | 41 | 1% |
| 13s | 34 | 1% |
| 20s | 30 | 1% |
| 7s | 19 | 1% |
| 45s | 18 | 1% |
| 5s | 17 | 1% |
What aspect ratios are used?
| Aspect ratio | Prompts | Share of stated |
|---|---|---|
| 16:9 | 821 | 51% |
| 9:16 | 629 | 39% |
| 1:1 | 78 | 5% |
| 4:3 | 23 | 1% |
| 2.39:1 | 17 | 1% |
| 2.35:1 | 13 | 1% |
| 3:4 | 10 | 1% |
| 21:9 | 8 | 0% |
| 4:5 | 6 | 0% |
| 2:3 | 1 | 0% |
| 19:6 | 1 | 0% |
Which categories are largest?
| Category | Prompts | Share of corpus |
|---|---|---|
| cinematic-realistic | 1,878 | 32% |
| fantasy-magical | 1,126 | 19% |
| sports-action | 1,001 | 17% |
| cinematic-showcase | 849 | 14% |
| brand-commercial | 755 | 13% |
| anime | 592 | 10% |
| short-film | 573 | 10% |
| cyberpunk-scifi | 569 | 10% |
| social-lifestyle | 556 | 9% |
| 3d-cartoon | 430 | 7% |
| person-character | 395 | 7% |
| animal-pet | 311 | 5% |
| music-video | 276 | 5% |
| food-drink | 274 | 5% |
| ugc-talking-head | 268 | 5% |
A prompt can carry up to three categories, so shares do not sum to 100%.
How was this measured?
Each prompt was ingested from its original public post and analysed into a fixed schema. Camera, lighting and aspect ratio were extracted with a controlled vocabulary; duration and the semantic fields (title, summary, subject, mood) were assigned by an LLM reading the full prompt text. Values outside the controlled vocabulary were rejected, so the term frequencies above reflect a closed list rather than free-text matching.
Corpus: 5,926 prompts, 1,196 authors, source posts from 2025-10-01 to 2026-09-10.