Orelon logoOrelon
Preise

Cinematography Practices That Make AI Video Look Cinematic

14. 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.

A practical guide to lens language, lighting, composition, movement, and color in AI video prompts, with workflows, examples, fixes, and decision criteria.

Most AI video that misses the mark does not fail because the model is weak. It fails because the shot was never described as a shot. A prompt like "a woman walking through a rainy city, cinematic" hands the generator a subject and a mood, but no lens, no light direction, no blocking, and no reason for the camera to be where it is. The output is usually clean, smooth, and emotionally flat.

Cinematography is a decision system. Every frame answers a small set of questions: where the camera sits, what it is looking at, what is lit and what falls into shadow, what sits between the lens and the subject, and what changes between the first second and the last. Models have no taste, but they are extremely responsive to decisions. Supply the decisions and the footage starts to feel directed.

This guide converts that craft into instructions a video model can actually follow, with prompt examples, a repeatable production workflow, the mistakes that cost the most time, and clear criteria for deciding where to spend effort.

Why camera language still decides whether AI footage works

Generation quality has improved faster than most people's shot planning. Motion is smoother, faces hold together longer, and lighting is more believable. That raises the bar rather than lowering it. When everyone can produce a moving image, the differentiator becomes whether the image means anything.

Three properties separate footage that reads as cinematic from footage that reads as generated:

  • Intent. The frame exists for a reason. A low angle makes a character dominant; a wide lens in a narrow corridor makes the space oppressive; a slow push-in implies dawning realization. If you cannot name the reason, the shot is decoration.
  • Constraint. Cinema is built from limits: a fixed lens set, one motivated light source, a locked camera. Constraints create style. Prompts that try to include everything produce mush.
  • Continuity. A sequence feels professional when light direction, palette, and lens character survive the cut. Single beautiful clips that do not match each other look like a demo reel, not a film.

Treat the model as a fast, literal crew. It will not protect you from a bad decision, and it will not invent one for you. Your job is to make the decisions small enough to describe and strong enough to see.

Write a shot brief before you write a prompt

The fastest way to improve output is to stop prompting and start briefing. A shot brief is five lines that fit on a sticky note.

The five-line brief

  1. Subject and wardrobe. Who or what, in specific terms — "a courier in a soaked navy rain shell," not "a person."
  2. Action. One physical verb with a beginning and an end — "lifts the crate onto the tailgate."
  3. Frame and lens. Shot size plus focal length — "medium-wide, 35mm, camera at chest height, four meters back."
  4. Light. Source, direction, quality — "single sodium streetlamp behind and above, hard rim, soft bounce from wet asphalt."
  5. Change. What moves during the clip — "the camera trucks right as the crate lands."

Here is the brief assembled into a prompt:

Medium-wide 35mm shot, camera at chest height four meters back. A courier in a soaked navy rain shell lifts a wooden crate onto a van tailgate, rain streaking past the lens. Single sodium streetlamp behind and above him creates a hard rim light; wet asphalt bounces soft orange fill into his face. Slow truck right as the crate lands, shallow depth of field, background traffic blurred into bokeh, cool blue shadows against warm highlights, fine grain, 24fps motion blur.

That prompt is long by text-to-image standards and short by film standards. It works because every clause removes an option rather than adding an adjective.

Use a consistent prompt order

Models weight the beginning of a prompt more heavily and drift when instructions arrive in random order. Pick an order and reuse it: subject → action → framing and lens → light → atmosphere → palette → motion → technical constraints. Consistency in prompt structure produces consistency in output, which is exactly what a sequence needs.

Cut the contradictions

Most weak prompts contain two or three instructions that cancel each other: "soft dreamy haze" plus "crisp high-contrast detail," or "handheld documentary energy" plus "locked-off symmetrical composition." Before generating, read the prompt and ask whether a real crew could execute all of it in one take. If not, split it into two shots.

Build a camera vocabulary the model can act on

Vague camera words are the biggest source of randomness. "Cinematic angle" means nothing; "low angle, 24mm, subject fills two-thirds of frame" means something specific and reproducible.

Focal length and distance

Focal length does two jobs at once: it sets field of view and it sets how space compresses. Naming a lens is often more effective than naming a shot size, because it implies both.

Intended effect Prompt phrasing What it does on screen
Environment and isolation "24mm wide, subject small in frame" Exaggerates depth, makes spaces feel vast or hostile
Natural observation "40mm, eye level, waist-up" Closest to human perspective, low drama, high clarity
Intimacy and compression "85mm, shallow depth of field" Flatters faces, blurs background into color
Detail and texture "100mm macro, water beading on metal" Turns surfaces into subject matter

Height and angle

Camera height is the most underused control in AI prompting, and it changes the emotional reading of a shot more than most style words. Eye level is neutral. Slightly below eye level gives quiet authority. A hard low angle turns a subject monumental or threatening. An overhead shot turns people into geometry and makes consequences feel inevitable. A dutch tilt should be used once, deliberately, and never as seasoning.

Movement verbs that survive generation

Movement is where AI video most often falls apart, so describe it in physical terms with a speed and a subject:

  • Dolly in / push in — camera moves toward the subject. Specify distance: "pushes in one meter over four seconds."
  • Truck / track — camera moves sideways, parallel to the subject.
  • Crane up / boom down — vertical movement that reveals scale.
  • Orbit — camera arcs around a static subject; keep the arc under 90 degrees for stability.
  • Handheld follow — camera trails the subject with small organic drift; add "subtle, walking pace" or it becomes shaky chaos.
  • Whip pan — fast rotation used as a transition; best kept to a single clip and cut on the blur.

Two rules keep motion believable. First, one action per clip. Second, short durations: three to five seconds per shot is the sweet spot for most engines, and you can extend later in the edit rather than asking for a ten-second continuous take that slowly degrades.

If you keep a personal prompt library of camera phrases that worked, you stop re-solving the same problems and start building a recognizable style.

Light like a gaffer, not like a search engine

Lighting is where amateur and professional AI footage diverge most visibly. "Good lighting" produces flat, evenly exposed images. Named sources with direction and quality produce shape.

Direction and quality first

Describe light the way a crew call would: source, direction, quality, ratio.

  • Source: window, streetlamp, practical table lamp, overcast sky, phone screen.
  • Direction: camera left, behind subject, three-quarters back, top-down.
  • Quality: hard (sharp-edged shadows) or soft (gradual falloff).
  • Ratio: how bright the lit side is compared with the shadow side — 2:1 is gentle, 8:1 is dramatic.

"Soft window light from camera left, subject's shadow side two stops down" is a sentence a cinematographer would recognize, and models respond to it because it describes a physical arrangement rather than a vibe.

Four reliable lighting recipes

  1. Golden hour rim. "Sun low behind subject, hard warm rim on hair and shoulders, soft bounce fill from warm pavement, lens flare minimal." Good for hopeful or nostalgic beats.
  2. Overcast interview. "Large soft source overhead and slightly camera right, no visible shadow edges, muted color, even skin tones." Good for documentary and confession.
  3. Neon night exterior. "Magenta practical sign camera left, cyan spill from a shop window camera right, wet ground reflecting both, subject half in shadow." Good for tension and urban isolation.
  4. Single-source chiaroscuro. "One hard lamp camera right, 8:1 ratio, background falling to near black, thin rim separating subject from darkness." Good for threat, secrecy, or grief.

Keep light consistent across shots

Continuity is a writing problem before it is a generating problem. Build a one-paragraph light bible for the scene — time of day, weather, key source, color temperature, contrast level — and paste the same phrasing into every prompt for that scene. If a wide shot says "late afternoon sun from camera left" and the close-up says "soft ambient light," the cut will feel wrong no matter how good each clip is individually.

Atmosphere, depth, and composition

Atmosphere is not decoration; it is a depth tool. Haze, fog, dust, smoke, and rain scatter light and separate foreground from background, which is precisely what makes an image read as three-dimensional rather than as a flat cutout.

Use atmosphere with restraint and with a stated density. "Light atmospheric haze catching a shaft of window light" gives volumetric beams and softened backgrounds. "Thick fog" erases your set. "Dust motes drifting in a sunbeam" adds texture on a close-up; the same phrase in a wide shot does nothing.

Build frames in layers

Every strong frame has at least two planes of interest. Name them explicitly in the prompt:

  • Foreground — out-of-focus railing, rain on glass, a shoulder in the near corner.
  • Midground — the subject doing the thing.
  • Background — context that comments on the subject.

A prompt that says "framed through a blurred doorway, subject midground at a kitchen table, cold blue window light behind her" gives the model a spatial problem to solve, and spatial problems produce depth.

Composition rules that survive generation

Rule-of-thirds placement still helps, but the phrasing matters. Instead of "rule of thirds," say "subject on the right third, looking into empty space on the left." That single change gives you both composition and eyeline, and eyeline is what makes a frame feel motivated.

Other reliable controls:

  • Negative space — "seventy percent empty sky above the subject" creates loneliness or scale.
  • Leading lines — "corridor lines converging on the subject" pulls the eye without effort.
  • Frame within frame — "subject seen through a car window" adds voyeurism and separation.
  • Centered symmetry — reserve for authority, ritual, or comedy; used everywhere it becomes a style crutch.

Aspect ratio and safe areas

Choose the frame before you generate. A 16:9 composition collapses badly when cropped to 9:16, because the subject that sat on the right third ends up half out of frame. If the deliverable is vertical, prompt vertical and compose with generous headroom. If you need both, generate the wider frame first and plan a center-weighted composition that survives a crop.

Color, contrast, and a consistent look

Color is the cheapest way to make a sequence feel like one film, and the easiest way to make it feel like a folder of clips from different creators.

Pick a palette before you pick a look

Name three colors and their relationship: warm amber against deep teal, desaturated sage with bone white, sodium orange against wet black. Then keep the wording identical across shots. Models interpret "teal and orange" inconsistently, but "cool teal shadows, warm amber skin highlights, muted greens in foliage" lands reliably.

Control contrast, not just saturation

Black level is a mood decision. Lifted blacks with low contrast read as memory, nostalgia, or haze. Crushed blacks with bright highlights read as thriller, heat, or confrontation. Say which one you want: "lifted blacks, low contrast, soft highlight rolloff" or "deep blacks, hard specular highlights, high micro-contrast."

Use film-emulation language carefully

A short, physical phrase does more than a brand name: "fine grain, gentle halation around highlights, slightly muted greens, subtle lens vignette." One or two of these per prompt is plenty. Stacking five emulation terms usually produces a muddy image that looks filtered rather than shot.

A repeatable workflow from board to final cut

Ad hoc generation is fine for experiments and terrible for deliverables. This workflow scales from a single scene to a short film.

Step 1: One-page treatment

Write the story in plain prose, then list the shots that are actually required. Most projects fail because they generate forty clips for a ninety-second piece and discover in the edit that three shots are missing.

Step 2: Look development with stills

Lock the visual language before spending time on motion. Generate stills for each key setup using an AI image generator, iterate on lens, light, and palette, and pick one frame per scene as your reference. Stills are cheap, fast to compare, and easy to share with a collaborator.

Step 3: Lock a look frame

Choose the single image that defines the scene. Write down the exact phrases that produced it — lens, light direction, ratio, palette, grain. That list becomes your continuity document for every subsequent prompt.

Step 4: Test clips before full batches

Generate two or three short variations per shot at low duration, review them side by side, and only then commit to a longer generation. Decide on a fixed review criterion: does the motion serve the action, does the light match the look frame, does the frame hold for the full duration. Judging against criteria beats judging against mood.

Step 5: Assemble and repair in the edit

The edit fixes more than generation does. Speed ramps hide awkward motion, a two-frame dissolve hides a continuity jump, stabilization rescues a handheld shot, and a grain pass unifies clips from different generations. Add sound early — room tone, footsteps, fabric — because audio does more for perceived realism than another hour of re-prompting.

Step 6: Finish and deliver

Deliver at the aspect ratio and length the platform expects. If you need several formats, compose for the widest and check the crop before you export. Reusable video templates can carry the technical setup so you spend your time on the decisions that actually change the image.

Mistakes that cost the most time, and how to fix them

Too many ideas in one clip. A prompt containing a costume change, a location change, and a camera move will produce a mess. Fix: one action, one location, one movement per clip, then join in the edit.

No stated light direction. The model defaults to flat frontal light. Fix: always name a source and a side.

Style words instead of physical description. "Epic" and "moody" are not instructions. Fix: replace each adjective with the physical thing that causes it — haze, low sun, hard shadow edge, wet ground.

Mismatched aspect ratios across a sequence. Fix: set the frame in the prompt and check every export.

Characters drifting between shots. Fix: describe wardrobe, hair, and build in identical wording every time, and generate a reference still to compare against.

Judging single clips instead of sequences. A shot that looks great alone can break a scene. Fix: review in context, in order, at final length, with sound.

Decision criteria: where to spend effort

  • Invest in look development first when the project is brand-facing, character-driven, or longer than thirty seconds. Consistency pays back immediately.
  • Invest in motion testing first when the scene depends on physical action — running, fighting, driving, dancing. Motion is the hardest thing to fix later.
  • Use image-to-video when the composition must be exact, such as a product hero frame or a match cut. Text-to-video is faster for exploration and for shots where the frame can drift.
  • Accept a simpler shot when the alternative requires a camera move the engine cannot hold. A locked-off frame with strong blocking beats a broken orbit every time.

FAQ

Do I need real cinematography knowledge to get good results? You need vocabulary more than experience. Learn four things — focal length, light direction, shot size, and one movement verb — and your output changes immediately. The rest can be learned by comparing your clips to frames you admire and naming the difference.

Why does my motion look wobbly or rubbery? Usually because the prompt asks for too much movement in too little time, or describes movement without a subject. Shorten the clip, reduce the number of actions to one, and add speed language such as "slow, steady push in over three seconds."

How do I keep a character consistent across shots? Write a fixed character string — age range, build, hair, wardrobe, distinguishing feature — and reuse it verbatim. Lock a reference image, keep the same lens and light phrasing, and avoid changing the time of day between shots in the same scene.

How long should each generated clip be? Three to five seconds per shot for most narrative work. Longer clips look impressive in isolation but tend to degrade in the middle, and they constrain your editing options. Generate short, cut tight.

Should I prompt the aspect ratio or crop afterward? Prompt it. Composition depends on the frame, and a crop made later usually destroys the framing you carefully built. If you need multiple formats, compose centrally and check each crop before delivery.

How many variations per shot should I generate? Three is a reasonable default: one safe, one with a different camera height, one with a different light ratio. Reviewing three options teaches you faster than generating thirty and hoping.

Put the craft into practice

The gap between AI footage that looks generated and AI footage that looks shot is not a secret model or a magic phrase. It is a brief, a lens, a light direction, a composition decision, and enough discipline to keep them consistent across a sequence.

Start with one scene. Write the five-line brief, lock a look frame, generate three short variations per shot, and cut them together with sound. Then repeat the same phrasing on the next scene and watch how quickly the work starts to feel like a film rather than a set of clips.

When you are ready to build the sequence, use the Orelon AI video generator to move from look frame to motion, keep your camera and lighting phrases in one place, and iterate in short, reviewable passes rather than long, hopeful ones.