Orelon logoOrelon
Preise

How to Analyze Movie Scenes for Better AI Video Prompts

16. Nov. 2026 · Von Orelon Team

KI-Video-Vorlagen entdecken

Lass dich von ein paar Community-Kreationen inspirieren und öffne dann eine Vorlage, um in Orelon weiterzuerschaffen.

Turn great film scenes into precise AI video prompts: a four-pass scene analysis method, prompt blocks, shot lists, continuity systems, and pacing rules.

Most people who complain that AI video looks generic are not using weak tools. They are writing weak directions. A prompt like "cinematic, moody, dramatic lighting" describes a feeling the model cannot verify, so it returns the average of everything it has seen — and averages look like stock footage.

Finished film scenes are the antidote. Not because you should copy them, but because every memorable scene is a stack of solved problems: where to put the camera, which way the light comes from, how long to hold before cutting. When you can read those decisions, you stop writing adjectives and start directing.

This guide is a working method for that translation — how to study a scene in four passes, how to record what you find, and how to turn it into instructions that hold up across a whole sequence rather than one lucky clip.

Why a finished scene teaches more than a style keyword

A style keyword is a compression of thousands of decisions into one word. That compression is convenient and almost useless as an instruction. "Gritty" could mean handheld camera, sodium streetlight, rain on a lens, or a color grade pulled toward green. The model picks one at random; you wanted another.

A scene study unpacks the compression. It gives you five categories of information that survive generation:

  • Geometry — where the subject sits, what overlaps them, how much empty space surrounds them.
  • Light logic — direction, hardness, color, and whether the source is visible in frame.
  • Optics — the implied focal length, depth of field, distortion, background compression.
  • Camera behavior — locked off, drifting, tracking, handheld, or craning, and at what speed.
  • Duration — how long the moment holds, which is a storytelling decision, not a technical one.

Keywords are seasoning. Those five categories are the meal. If you can name them for a scene, you can rebuild that scene's effect in a completely different setting — which is the whole skill.

There is a second benefit that rarely gets mentioned: analysis gives you a shared vocabulary with collaborators. "Push in slowly from a medium shot, single cool key from the left, two stops under on the fill" is a note a cinematographer, an editor, and an AI tool can all act on. "Make it more epic" is not.

The scene autopsy: four passes, one notebook

Watch the same scene four times, each pass dedicated to one layer. It feels painfully slow the first time and automatic by the fifth scene. Keep a notebook or a plain text file open; the writing is what trains the eye.

Pass one: geometry and framing

Pause on three stills — beginning, middle, end of the shot. For each, answer four questions in plain language:

  1. Where is the subject in the frame — center, left third, right third, lower corner?
  2. What occupies the foreground, and does it overlap the subject?
  3. Where is the negative space, and what emotion does that empty area create? Space above a head reads as vulnerability or stillness; space beside a figure reads as isolation.
  4. What is the shot size — extreme wide, wide, medium, close, insert?

Center framing tends to read as confrontation or ceremony. Off-center framing reads as observation. A tight frame with clipped headroom creates pressure. None of these are rules; they are tendencies you can choose deliberately once you can name them.

Pass two: light sources and ratios

Count the light sources you can see or reasonably infer: a window, a practical lamp, a phone screen, a streetlight, fire, a doorway to a brighter room. Then judge the contrast ratio between the lit side and the shadow side.

  • Nearly black shadows with a hard key: drama, menace, secrecy.
  • Soft, filled shadows: safety, memory, warmth, comedy.
  • A bright rim separating the subject from a dark background: importance, separation, occasionally cliché if overused.

Then note color separately. A warm key against a cool background is the most common tension palette in contemporary film. A single saturated object in an otherwise muted frame becomes a visual magnet, and audiences will look exactly where you put it.

Finally, ask whether the light is motivated. If there is a lamp in frame, the light should plausibly come from it. Unmotivated light is not wrong — it is a stylization choice — but mixing motivated and unmotivated sources in the same shot is where prompts turn to mud.

Pass three: lens and camera behavior

Ask what the lens is doing to space. Wide lenses stretch rooms and make closeness uncomfortable. Long lenses compress background into a wall behind the face and isolate the subject from their world. Shallow depth of field says this is the only thing that matters here. Deep focus says everything in this room is part of the problem.

Then describe the movement in one sentence. If you cannot, the shot probably does not have a clear intention. A slow push-in builds realization. A lateral track builds momentum. Handheld builds immediacy and unease. A locked-off frame builds formality and hands the scene to the performance. Speed matters too: a push that takes eight seconds feels like dawning understanding; the same push in two seconds feels like an attack.

Pass four: duration and cut rhythm

Watch with a mental stopwatch. How long does each shot hold? What does the cut land on — a face, a hand, an empty room, a doorway? Shortening shots accelerate; holding one shot well past comfort produces dread or grief. Notice also what the scene withholds: the reaction shot it refuses to give you is often the reason the moment stays with you.

Editing rhythm is the hardest layer to prompt directly, because generators produce clips rather than cuts. You reproduce rhythm by generating shots of planned lengths and assembling them with intent. That means the analysis has to end up on a timeline, not just in a prompt field.

Turning observations into prompt blocks

Free-form prompts encourage free-form results. A block structure keeps every decision visible and editable, so when something misses you can diagnose which line failed.

The seven-block skeleton

  1. Subject and wardrobe — who, age range, build, clothing, one distinguishing detail.
  2. Shot and framing — shot size, subject position, foreground elements, negative space.
  3. Lens and depth — wide/medium/long feel, background blur, distortion notes.
  4. Lighting — source count, direction, hardness, contrast ratio, motivated or stylized.
  5. Palette and texture — dominant hues, one accent, grain or cleanliness.
  6. Camera and performance — movement, speed, what the subject does and does not do.
  7. Constraints — what must not appear, no text overlays, no extra people.

Worked example: the quiet diner

Before: "Lonely woman in a diner at night, cinematic, moody lighting, film grain, highly detailed."

After: Subject: woman, late thirties, dark hair pulled back, grey coat over a work shirt, no jewelry. Shot: medium, subject seated at the counter on the right third, window reflection overlapping her shoulder as foreground, wide empty counter filling the left of frame. Lens: 50mm feel, background booths softly out of focus. Light: single warm tungsten pendant above the counter as key from upper left, weak cool spill from the window as fill, shadow side about two stops down. Palette: amber and deep teal, low saturation apart from the key. Camera: locked off, no movement. Performance: she stirs a cup once, looks toward the window, does not smile. No other customers, no visible signage text.

The second version is longer, but every sentence is a decision you can defend, reuse, or fix. That is the difference between hoping and directing.

A useful habit: keep a personal file of block skeletons that produced good results, and swap only the subject and location lines. You will build a small library of reliable lighting and camera recipes that work across projects — and you can browse the prompt library when you want a structural reference instead of a blank page.

Shot lists that behave like a story

A single beautiful clip is a demo. A sequence is a film. Before generating anything, write a plan where every shot has a distinct job.

One job per shot

A practical six-shot spine that works in almost any genre:

  1. Establishing — where are we, what time, what emotional weather.
  2. Approach — the subject enters, or the camera closes distance.
  3. Reaction — the face that tells us what the situation means.
  4. Detail insert — a hand, a clock, a cup, a door handle. This is the shot that buys believability.
  5. Turn — something changes: light, distance, direction of movement, who is in frame.
  6. Release — the shot that lets the audience breathe, or deliberately refuses to.

If two shots share the same job at the same scale, cut one. Sequences feel thin when every shot carries the same information at the same distance.

The continuity ledger

Keep a second short document listing facts that must not change: time of day, weather, which side the key light comes from, what the character is wearing, whether a door is open, which direction traffic moves. Check the ledger before each generation, not after. Most continuity errors are not model failures; they are unstated assumptions.

Keeping a character and a place stable

Anyone can produce one good frame. Holding a face, a room, and a look across eight shots is the actual craft.

Describe people in fixed paragraphs

Write five lines — age range, build, hair, wardrobe, one distinguishing feature — and paste them verbatim into every prompt for that character. Do not paraphrase between shots; paraphrase is where drift enters. Generate a reference still in the AI image generator and keep it beside your prompts while you work.

Crucially, also keep the lighting family consistent. Identity reads as broken when light direction flips between shots far more often than when a face subtly changes.

Describe places like a location scout

List architecture, materials, dominant colors, weather, time of day, and the position of the primary light source. If two shots happen at the same hour, shadows should fall the same way. Audiences rarely notice when this is right and feel it immediately when it is wrong.

Freeze one style sentence

Write a single sentence covering contrast level, grain, palette, and overall finish. Paste it unchanged into every prompt in the sequence. Vary the shot, never the style sentence. It is the cheapest consistency upgrade available, and it makes your sequence look intentional even when individual clips are imperfect.

Pacing: emotion is a duration decision

Rhythm lives on the timeline, not in the prompt. Useful starting points: two to three seconds for urgency, four to six for observation, eight to ten for dread or grief. A chase cut from four-second shots feels sluggish; a funeral cut from two-second shots feels like a highlight reel.

Try an arc across five clips:

Shot Length Framing Movement Purpose
1 6s wide static establish distance
2 4s medium slow push close the gap
3 2s close handheld break composure
4 3s insert static deliver the information
5 9s wide static return to distance, changed

That structure works across genres because it is about information control rather than spectacle. Generate each clip slightly longer than you plan to use, then trim on the timeline where you can actually feel the pacing. Sound design — room tone, a single impact, a low bed — fixes rhythm faster than any re-render.

A practice loop you can run in an afternoon

  1. Pick three scenes with different jobs — one dialogue scene, one silent physical beat, one transition or montage. Variety of reference teaches more than three versions of the same look.
  2. Run the four passes on each and write your findings in plain language, no jargon required.
  3. Extract only what transfers — geometry, light logic, camera behavior, duration — and discard everything that only worked because of the original story.
  4. Rebuild one scene in a new world: change era, wardrobe, palette, and location, keep the structural lesson.
  5. Generate the hardest shot first, four low-commitment variants. Solving the difficult shot early prevents a polished sequence with one broken link.
  6. Freeze the style sentence and character paragraph before generating the rest.
  7. Assemble with planned durations, add sound, then re-run only the weakest shot. Do not regenerate the entire sequence to fix one clip.

Steps five through seven are where the AI video generator does its work, but the leverage is in steps one through four. Most people skip straight to generation and wonder why the result feels like a mood board instead of a scene.

Mistakes that make film references backfire

  • Reproducing a famous shot frame for frame. It reads as derivative and can create real legal exposure. Take the principle, change the surface.
  • Adjective stacking. Ten mood words dilute each other. Three precise technical choices beat ten vibes.
  • Mixed lighting logic. A hard key plus glowing ambient fill plus a rim light from nowhere produces mush. One logic per shot.
  • Ignoring screen direction. If a character exits left, having them enter from the left in the next shot disorients the viewer without any obvious reason why.
  • Cramming two ideas into one six-second clip. One beat per clip. Two beats need two clips.
  • Judging output by attractiveness instead of by the shot list. Compare against written intent, or you will keep a pretty accident and lose the plan.
  • Changing the style sentence mid-sequence. It is the single most common cause of a sequence that looks assembled from different projects.
  • Generating before planning. Ten minutes with a notebook saves an hour of re-rolls.

Influence versus imitation: where to draw the line

A workable rule: take principles freely, take surfaces carefully.

Principles — contrast ratios, framing logic, cut rhythm, use of negative space, the decision to hold a shot — are shared craft knowledge built over a century of filmmaking. Learning them is not theft; it is literacy.

Surfaces — a recognizable actor's likeness, a protected character, a signature costume, a living director's unmistakable visual signature recreated for commercial work — are where problems begin. In practice: change setting, wardrobe, and palette; keep the structural lesson. Your result should be unrecognizable as a copy and obvious as a craft decision.

Decision criteria: analyze more, or generate more?

Use this quick test when you feel stuck:

  • Outputs look different from each other in style → your style sentence is missing or drifting. Fix the prompt system, not the model.
  • Outputs look consistent but boring → you are repeating the same shot size and light logic. Analyze a scene with a different job.
  • One shot keeps failing → break it into two simpler shots rather than adding more adjectives.
  • The sequence feels slow → check durations, not content. Trim two seconds before rewriting anything.
  • The sequence feels confusing → check screen direction, time of day, and which side the key light comes from.
  • Everything looks fine but nothing lands emotionally → you have no reaction shot. Add a face at the right moment.

If more than two of these apply, spend an hour on analysis before spending another hour on generation. If none apply, stop studying and start rendering — analysis without output is just a hobby. Browsing finished work on the Orelon blog is a fast way to calibrate what "finished" looks like before you judge your own clips.

FAQ

How many scenes do I need to study before I improve?

Ten scenes analyzed properly will change your prompting more than a hundred watched casually. Four passes each, written notes, one rebuild exercise. The writing is the part that trains the eye.

Can AI video tools actually follow detailed cinematography prompts?

They follow concrete, constrained language far better than mood words. Framing, light direction, contrast, and camera behavior are reliably steerable. Exact lens millimeters and f-stops are approximations — describe the visible effect instead, such as "background heavily blurred, face isolated."

How do I keep a character consistent across many clips?

Three anchors: a fixed descriptive paragraph pasted verbatim, a reference still, and a consistent lighting direction. Most identity drift comes from lighting and framing changes rather than from the model forgetting a face.

Should I write one long prompt or several short ones?

One prompt per shot, structured in blocks. Long multi-beat prompts produce clips that rush through early beats and ignore later ones. Two ideas mean two clips.

What if a clip looks great but breaks the sequence?

That is the most common trap. Judge against the shot list, not against how attractive the frame is. A gorgeous shot that breaks continuity costs more than a plain shot that holds the sequence together.

How long should each generated clip be?

Generate a little longer than you plan to use, then trim in the edit. Handles give you room for transitions and let you tune pacing on the timeline where you can feel it.

Do I need film school vocabulary to do this?

No. "Light comes from the window on the left, shadows are dark, camera does not move" is a complete and effective direction. Precision matters; jargon does not.

Start with one scene, one shot, one decision

The fastest way to improve AI video is not a newer model. It is a sharper eye. Choose a scene you love, watch it four times, and write down its geometry, light, lens, and rhythm. Then rebuild that structure somewhere else entirely, with your own character, your own palette, and your own ending.

Orelon is built for exactly this kind of work — cinematic ideas in motion, where a written decision becomes a moving image you can refine shot by shot. Open the video generator, bring your shot list and your style sentence, and direct the scene instead of hoping for it.