Orelon logoOrelon
料金

Cinematic Composition and Lighting: A Practical AI Video Guide

2026年11月16日 · Orelon Team 著

AI動画テンプレートを見る

着想のためにコミュニティ作品をいくつか閲覧し、任意のテンプレートを開いて Orelon で作成を続けましょう。

Learn how framing, key and fill light, rim light, and color grading shape mood, and how to translate each technique into reliable AI video prompts.

A frame can be lit beautifully and still say nothing. What separates a shot that lingers in memory from one that scrolls past is rarely the subject matter — it is the arrangement of space, the direction of light, and the discipline of the color palette. Those three decisions do more emotional work than most dialogue.

This guide breaks down cinematic composition and lighting as practical craft, then shows how to translate each technique into prompts and settings for AI video generation. No abstract theory without an application: every section ends with something you can test in your next render.

Why composition and lighting outperform spectacle

Audiences do not consciously read a frame. They feel it. A subject pushed to the left edge with empty space ahead reads as anticipation; the same subject centered and lit from below reads as threat. Nothing in the story changed — only the geometry and the light.

That is why studying high-impact scenes is useful even if you never plan to recreate them. You are not copying a look. You are learning a grammar: which visual choices reliably produce which emotional responses.

For AI video work this matters more than in traditional filmmaking, because you cannot adjust a real set. You can only describe. The more precisely you understand composition and lighting, the more precisely you can describe — and the fewer renders you burn hunting for a look you could have specified in the first place.

Decide the shot before you open the tool

Most weak AI video output starts with a prompt, not a plan. Reverse that order.

The five decisions that come before any tool

Write these down in plain language before you touch a generator:

  1. Subject placement — center, thirds, or edge-of-frame.
  2. Camera height — eye level, low, or high angle.
  3. Lens feel — wide and immersive, normal and neutral, or long and compressed.
  4. Lighting direction — where the key comes from in clock terms relative to the face.
  5. Contrast ratio — soft and open, or hard and secretive.

Five lines. That is your shot plan. Everything after this is execution: writing the prompt, choosing a reference image, generating variations.

If you are working from a mood board, pair the board with a written shot plan. Images communicate taste; the written plan communicates intent. AI tools respond much better to intent.

Framing: controlling where the eye lands first

Composition is attention management. The viewer's eye will enter the frame somewhere, travel along something, and rest somewhere. Your job is to make those three movements deliberate.

Rule of thirds, and when to abandon it

Placing a face on a vertical third line gives the frame breathing room and creates a natural relationship between subject and environment. It is the default for a reason: it feels balanced without feeling inert.

But the rule of thirds is a starting point, not a law. Two deliberate exceptions:

  • Center framing creates confrontation, symmetry, or unease. Use it when you want the viewer to feel addressed directly — a stare, a ritual, a formal portrait.
  • Extreme edge placement with large negative space creates isolation or anticipation. The subject has somewhere to go, or somewhere they are avoiding.

When you prompt, say the placement explicitly. "Subject on the left third, looking into empty space on the right" produces a different image from "portrait of a woman in a room."

Leading lines and depth cues

Depth is what stops an image looking like a sticker on a background. Three cheap ways to create it in a generated frame:

  • Architectural lines — corridors, window frames, railings, rows of lights receding toward the subject.
  • Foreground occlusion — something slightly out of focus between camera and subject. A doorway edge, a plant, a shoulder.
  • Atmospheric separation — haze, dust, or smoke that lightens as it recedes, mimicking aerial perspective.

All three can be requested in a sentence. "Foreground doorway edge softly out of focus, corridor lights receding toward subject, light haze" will noticeably change your output and cost you nothing.

Negative space as emotional pressure

Empty space is not wasted space. A subject occupying twenty percent of the frame with eighty percent emptiness communicates something specific: smallness, waiting, power imbalance, or contemplation.

Use negative space when the story is internal. Use full frames when the story is external. A crowded frame with no empty area reads as pressure or claustrophobia — useful, but exhausting if sustained across a whole sequence.

A practical rule for sequences: alternate. Two tight, busy frames followed by one wide, empty frame lands harder than three busy frames in a row.

Lighting: building a face out of three sources

Lighting has a vocabulary, and it is smaller than people expect. Learn key, fill, and rim, plus practicals, and you can describe almost anything.

Key light: direction sets the meaning

Where the primary light sits relative to the subject determines what the audience reads about them.

  • Frontal key — flattens shadows, feels open and commercial, hides subtext.
  • 45-degree key — the classic portrait position; models the face, keeps both eyes readable.
  • Side key (90 degrees) — splits the face into lit and unlit halves. Ambiguity, duality, inner conflict.
  • Backlight-dominant — subject goes semi-silhouette. Mystery, or the sense that something is behind them.
  • Underlight — reverses natural expectation. Instantly unsettling; use sparingly.

Direction is more powerful than intensity. Changing brightness by thirty percent rarely changes the meaning; moving the key ninety degrees always does.

Fill light: the contrast ratio dial

Fill is not about brightness — it is about how much shadow detail survives. That is the contrast ratio.

  • 1:1 — no visible shadow. Flat, clinical, or comedic.
  • 2:1 — gentle modelling. Natural, flattering, documentary.
  • 4:1 — clearly divided face. Drama, weight, intent.
  • 8:1 and above — near-black shadows. Noir, secrecy, threat.

In AI prompts, describe the ratio in plain terms: "soft fill keeping shadow detail visible" versus "deep unlit shadows on the right side of the face." Vague words like "moody" get interpreted inconsistently; ratio language is far more reliable.

Rim light and practicals: separating subject from background

A rim light traces an edge of the subject with a thin bright line. Its job is separation — telling the eye where the subject ends and the background begins. Without it, subjects merge into dark backgrounds and the frame reads muddy.

Practicals are in-frame light sources: a lamp, a window, a screen, a streetlight. They do two jobs at once — they motivate the lighting logically, and they give you controllable bright points to compose around.

Prompts that mention a visible practical almost always produce better spatial logic than prompts that describe light from nowhere. "Lit by a single desk lamp at frame left" is a complete lighting plan.

Color: temperature, palette, and the grade

Color is the fastest emotional shortcut available, and the easiest to overdo.

Warm and cool as psychological framing

Warm tones (amber, ochre, tungsten) read as intimate, nostalgic, safe, or artificial. Cool tones (steel, cyan, moonlight) read as distant, clinical, isolated, or truthful.

The most useful pattern is not "warm scene" or "cool scene" but a mix: warm subject against cool environment, or the reverse. That contrast creates a hierarchy. The viewer knows instantly what the frame cares about.

Build a palette you can repeat

Pick three anchor colors and hold them across a sequence: a dominant, a secondary, and one accent reserved for the emotional beat. Write them into your prompt library as named palettes — "amber, slate, and pale gold" — rather than re-describing them each time.

Repeatable palettes are how sequences feel intentional. If shot one is teal and amber, shot five should still be recognizably teal and amber, even if the location changed.

A contrast ratio cheat sheet

Ratio Look Best for
1:1 Flat, shadowless Comedy, clinical, corporate
2:1 Soft modelling Interviews, natural drama
4:1 Divided face Conflict, decision scenes
8:1 Near-silhouette Noir, threat, secrecy
Silhouette Shape only Reveals, mystery, scale

Keep this table next to your shot plan. Whenever a render feels emotionally wrong, check the ratio before you rewrite the entire prompt.

Translating film language into AI prompts

An AI video generator does not know what "cinematic" means. It knows what patterns co-occur with words. Your job is to supply words with strong, consistent visual consequences.

A prompt skeleton that works

Use a fixed order so you can isolate variables when something breaks:

Subject → placement → camera height and lens → lighting setup → palette → atmosphere → motion

Example: A man in a wool coat, placed on the right third facing left, eye-level medium shot on a 50mm lens, single key from frame left at 45 degrees with soft fill, amber and slate palette, light fog, slow push-in.

That sentence contains seven independent decisions. If the result is wrong, you can change one and re-render. That is the entire advantage of structured prompting.

Words that do nothing versus words that do work

"Beautiful," "stunning," and "masterpiece" are noise. They do not map to a lighting direction or a lens. Replace them with physical descriptions:

  • Instead of "dramatic lighting" → "hard key from above right, deep shadows on the left cheek"
  • Instead of "shallow depth of field" → "background bokeh, subject sharp at two meters"
  • Instead of "moody" → "cool ambient with a single warm practical at frame right"

If you want a head start, browse the prompt library and notice how the strongest entries are almost entirely physical description.

A worked example: three shots, one scene

Take a simple premise — a person waiting in a hotel lobby at night. Three different emotional readings, same location:

Shot A, anxious. Subject on the left third, wide frame, large negative space to the right. Cool ambient at 2:1, one distant practical. Motion: static camera, subject slightly still.

Shot B, powerful. Center framed, low angle, hard key from frame right at 8:1. Warm practical behind for a rim. Motion: very slow push-in.

Shot C, isolated. Subject small in frame, long lens compression, silhouette against a bright doorway. Motion: none, just ambient movement in the background.

Same room, same character, three meanings — produced entirely by composition, light, and lens choice. That is the whole craft in one exercise, and it is also a good test of any AI video generator: can it hold a location while you change the meaning?

Mistakes that flatten a frame

  • Two equal light sources. No hierarchy means no focal point. One source should clearly dominate.
  • Subject touching the frame edge unintentionally. Either commit to edge framing or give clearance.
  • No separation from the background. Add a rim, a practical, or a value difference.
  • Palette drift across shots. Fix anchor colors and reuse them.
  • Describing mood instead of physics. Mood words produce random results; physical words produce repeatable ones.
  • Overloading one prompt with five ideas. One shot, one intention.

A repeatable workflow

  1. Write the shot plan (five decisions).
  2. Generate three or four still frames first using an AI image generator. Iterate on composition cheaply before spending motion renders.
  3. Lock the frame, then add camera movement.
  4. Generate two or three motion variations of the locked frame.
  5. Assemble, then check palette and contrast continuity across the sequence.
  6. Only then refine detail: grain, diffusion, grade.

Iterating on stills first is the single biggest time saver. Composition errors are cheap to fix in a still and expensive to fix in a moving clip.

FAQ

Do I need to know film theory to get good AI video results? No, but you need to know the physical vocabulary: light direction, contrast ratio, lens feel, and subject placement. Those four concepts cover most of what theory would otherwise teach you.

How many lighting details should one prompt contain? Two or three. A key direction, a fill or ratio note, and a practical or rim. More than that and the model starts averaging your instructions.

Why do my frames look flat? Usually one of three causes: frontal key light, no background separation, or a palette with no hierarchy. Check those in that order.

Should I use the same palette for an entire project? Yes, with one accent that shifts by scene. Consistency is what makes separate clips feel like one film.

How do I make a subject look powerful without changing the story? Low angle, center framing, hard key from one side, and a slow push-in. Four changes, no script edits.

What is the fastest way to test a look? Generate the same prompt with three different lighting directions at low resolution. Compare side by side, pick one, then scale up.

Start building frames that mean something

Composition and lighting are not decoration added after the idea. They are how the idea becomes visible. Once you can specify placement, key direction, contrast ratio, and palette in plain language, AI video stops being a slot machine and starts being a camera you can point.

Take one scene you already have in mind. Write the five decisions, generate four stills with different lighting directions, then animate the winner. You can start right now with Orelon — and if you want to see how structured prompts translate into movement, the Seedance examples are a good place to compare notes.