Orelon logoOrelon
价格

Shot Composition and Lighting for AI Video: A Working Guide

2026年11月14日 · 作者:Orelon Team

探索 AI 视频模板

浏览社区创作获取灵感,打开任意模板即可在 Orelon 中继续创作。

Learn how to turn framing, lens depth, and lighting direction into precise AI video prompts that hold a scene together from the first shot to the last.

Most generated video does not fail because the model is weak. It fails because the prompt described a subject and left the camera and the light to chance. A model will happily invent a face, a street, and a colour palette. What it will not invent is your point of view.

Composition and lighting are the two levers that turn a rendering into a shot. They are also the two things people skip, because "cinematic" feels like a mood you describe rather than a set of decisions you make. It is not a mood. It is a camera position, a lens, a subject placement, and a light source with a direction. This guide turns those decisions into language you can paste into a prompt and reuse across a scene.

Why a Shot Reads as Cinema Before Anything Moves

Audiences do not analyse coverage. They read signals. A face lit from one side, with a bright rim separating it from a dark background, reads as someone with something to hide. The same face under flat, even illumination reads as someone in a tutorial. Geometry and light carry the story before a single line of dialogue arrives.

In generated video this matters more than in a traditional shoot, because there is no gaffer, no continuity supervisor, and no operator framing the next setup. There is your text, a reference image or two, and a training distribution. Vague words produce a coherent but generic frame: technically clean, emotionally anonymous.

The practical argument is speed. Iteration on generated clips is cheap, but only if you iterate on purpose. A team that specifies a shot usually lands it in two or three passes. A team that types "cinematic close-up of a woman in a city" burns a dozen attempts and still cannot match the result to the next shot. Precision is not fussiness. It is the shortest route to a usable take.

Composition: The Decisions That Actually Change the Image

Composition is the cheapest variable to control and the most visible when it goes wrong. Four choices do most of the work.

Choose the delivery ratio before you generate

Decide the format first. A 16:9 frame leaves room for landscape, two-shots, and lateral movement. A 9:16 vertical frame wants a single subject, centred or placed low, with air above the head. Generate in the wrong ratio and you either crop away the composition you designed or letterbox it into a compromise.

Also reserve the zone where captions or titles will live. A subject placed dead centre in a vertical frame leaves nowhere for text, and retrofitting a crop later will slice into hands, faces, and gestures. If the clip ends up in a feed with a caption band across the lower third, compose for that band from the start.

Place the subject with intent

Third-based framing is shorthand, not law. What matters is imbalance with purpose: put the subject off-centre and let the empty side carry meaning — threat, possibility, distance, loneliness. Centred framing reads as formal or confrontational, which suits an address to camera or a deliberate punchline.

Say it explicitly. "Subject on the left third, facing frame right, empty corridor on the right" is worth far more than "well composed." Add the facing direction too; a subject looking into empty space feels stuck, while a subject looking toward the open side of the frame feels like they are heading somewhere.

Build three depth planes

Flatness is the most common tell of amateur generated footage. Fix it by naming two planes, ideally three: a foreground element (a doorway edge, blurred leaves, a shoulder, rain on glass), the midground subject and action, and a background of architecture, weather, or crowd.

Specify a foreground element and the model generates parallax, occlusion, and the sense of a camera standing somewhere real. Add "slight foreground blur" for a depth cue without softening the face. If the background is busy, name what should fall out of focus instead of asking for "shallow depth of field" and hoping.

Use lens language as an instruction

Focal length is a compositional command. Wide lenses exaggerate space and proximity; long lenses compress backgrounds and isolate. In prompt form: "24mm wide angle, exaggerated perspective, low camera position" reads as scale and unease; "50mm, natural perspective" reads as dialogue; "85mm, compressed background, subject isolated" reads as intimacy.

Naming a lens also nudges distortion and depth of field predictably, which is how several separate shots end up feeling like the same camera package. Pick two lenses per project and stay with them.

Compose the Move, Not Just the Still

A still frame is composed once. A moving shot is composed along a path. Decide whether the camera is static, panning, tracking, or handheld, and where the subject sits at the start and end of the move. "Slow dolly in, subject stays on the left third" gives you a shot whose composition survives the whole clip. "Dynamic camera" gives you drift.

Movement should have a reason. A push-in builds attention; a lateral track reveals; a handheld drift suggests instability. If none of those is true for the beat, lock the frame. A static shot with good light beats a wobbling push-in almost every time, and it is far easier to cut against.

Lighting Described in Plain Language

Lighting is where prompts collapse into adjectives: moody, dramatic, glowing. Those words do something, but they never say where the light comes from — and direction is what creates shape on a face. You do not need a lighting diagram in the text. You need three facts: the direction of the key, its quality, and how much the shadow side is lifted.

Direction comes first

"Key light from camera left" produces a different face than "key light from above." Front light flattens and reassures. Side light reveals texture and ambiguity. Backlight turns a subject into a silhouette and makes atmosphere visible. Underlight unsettles. State where the source is, and when it matters, what it is: window, neon sign, phone screen, overhead fluorescent, firelight.

A useful habit is to describe the shadow rather than the light. "Shadow falling across the right cheek" is unambiguous; "dramatic lighting" is not. Shadows tell the model where the source sits and how far it falls off.

Hard or soft, and how much fill

Quality describes hardness. A hard key creates crisp shadow edges and reads as midday, bare bulbs, or spotlights. A soft key wraps, flatters, and reads as overcast sky or a large window. Ratio describes how much the shadow side is lifted: high contrast feels tense, low contrast feels commercial and safe.

Useful phrasing: "single soft key from a large window camera right, minimal fill, shadows falling left," or "hard low key from a practical lamp, deep falloff, background almost black." Both sentences contain a source, a direction, and a ratio — which is why they work.

Motivated light versus designed light

Motivated light has a visible or implied source inside the scene: a lamp, a doorway, a screen, headlights. It makes an image feel observed rather than lit, and models handle it well because it appears constantly in reference footage. Designed light ignores realism for effect: coloured gels, a shaft through haze, a glowing horizon behind a figure.

Both work. Pick one per shot. Mixing them produces the strange half-real look where a face is lit by nothing in particular, and the audience cannot tell whether they are in a real room or a stylised one.

Colour temperature as narrative information

Colour is a mood shortcut models understand extremely well. Warm amber reads as memory, safety, or dusk; cool blue reads as isolation, night, technology. Mixed temperatures inside one frame — warm face, cool window — create tension and depth without extra geometry.

Describe colour as a property of the light rather than a filter. "Warm tungsten practicals, cool moonlight through the window" gives the model a reason for the contrast. "Teal and orange grade" invites a flat overlay that swamps skin tones and makes every shot look like the same advertisement.

Turning Theory into Shot Cards

Theory only helps if it survives contact with a generation queue. A shot list says "Shot 4 — Anna in corridor, close-up." A shot card says:

Shot 4 — Anna, medium close-up, 85mm, subject on left third facing right, corridor receding behind with two visible doorways. Key light from a window camera left, hard quality, deep shadow on the right side of her face, thin rim from a distant exit sign. Cool ambient, warm spill on the cheek. Static camera.

That card can be generated, compared with the next one, and revised. It also makes continuity obvious: if shot 3 was a 24mm wide with flat overcast light, you now know exactly why the cut feels wrong.

Build the card in four lines and keep the order fixed: subject and wardrobe, framing and lens, light, then camera movement. A fixed order is not bureaucracy — it stops you from forgetting the light on the shot where it matters most.

Iterate one variable at a time. Fix framing and placement with simple light first. Once the geometry holds, add lighting detail, then atmosphere and movement. Changing everything at once means you cannot tell which phrase broke the shot, and you will end up reverting a change that was actually helping.

A Worked Example: Three Shots on a Night Platform

Take a scene with no dialogue: a woman waits on a train platform at night, then leaves.

Shot 1, establishing. "Wide shot, 24mm, night platform, woman small in frame on the right third, long empty platform receding left, overhead fluorescent strips creating pools of hard light with deep shadow between them, slight haze, static camera." The composition gives her smallness against architecture, and the hard overheads create rhythm and depth.

Shot 2, close-up. "Medium close-up, 85mm, same woman, key light from an overhead fluorescent above and slightly right, hard shadow under the brow, cool ambient fill from the platform, background compressed and out of focus with a single warm vending machine glow." Same source logic at a different scale, so the cut holds — the light has a reason in both frames.

Shot 3, departure. "Medium wide, 35mm, tracking behind the subject as she walks toward the exit, silhouette against a bright doorway, hard backlight, foreground pillar passing through frame." The bright exit becomes the motivated source, and the pillar supplies the parallax that sells the movement.

Three shots, one light logic, one lens family, one colour story. That is what a scene looks like when composition and light are decided before generation instead of salvaged afterwards.

Holding a Look Across a Whole Sequence

Consistency is where ambitious projects fail, and the fix is to reduce variables and reuse anchors.

Lock the character with a clear reference image plus a written description pasted into every card: age range, hair, wardrobe, distinguishing features, resting expression. A reference image alone drifts in styling; a written description alone drifts in likeness. Together they hold.

Lock the light logic in one sentence you repeat across the scene — "single hard source from camera left, cool ambient, no fill" — and change only the source's size and distance as the shot size changes. Lock the lens family for the same reason.

If a shot still drifts, do not regenerate blindly. Compare it with the anchor card and identify which element moved. Usually it is wardrobe detail, hair, or the direction of the key. Naming the drift is faster than re-rolling, and it teaches you which phrase in your card was doing the real work.

Build a contact sheet of every generated still and ask three questions before judging beauty: does the light direction stay consistent within the scene, does the subject hold screen position across the cut, and does the lens feel like the same package? Beauty is negotiable. Discontinuity is not.

Fix It in the Prompt or Fix It in the Edit

Not every problem deserves another generation. Structural problems go back to the prompt; cosmetic problems stay in the timeline.

Symptom Better fixed by
Wrong framing or subject placement Prompt — regenerate with explicit placement
Slight contrast or colour mismatch Edit — match with a grade
Face drifting from the reference Prompt — tighten description, re-anchor
Background too busy Prompt — name what should fall out of focus
Clip a beat too long or short Edit — trim or retime
Light direction contradicting the scene Prompt — restate the source, louder

A useful threshold: if fixing it would require adding or moving something physical inside the frame, it belongs in the prompt. If it only requires changing tone, timing, or scale, the edit will be faster and cheaper.

Mistakes That Give AI Footage Away

  • No stated light direction. The model picks a flattering default and every shot looks like the same stock image.
  • Too many sources. Three named lights plus atmosphere plus practicals turns into mush. One key, one accent.
  • Centred subject with no reason. Formal framing is powerful only when it is a choice.
  • Empty foreground. No near-field element means no depth, and no depth means no cinema.
  • Movement as default. "Dynamic camera" produces drift. Specify the move or lock the frame.
  • Adjectives without nouns. "Epic lighting" is not a description. "Hard backlight through haze" is.
  • Ignoring the delivery ratio. Compositions designed for widescreen fall apart in vertical crops.
  • Rewriting everything at once. Ten changed variables and no idea which one fixed the shot.

FAQ

Do I need film-school vocabulary to get better shots? No. Three facts carry most of the weight: direction ("from camera left"), quality ("hard" or "soft"), and source ("window," "neon," "overhead fluorescent"). Add a lens and a subject position and you already out-describe most prompts.

Why does my subject keep drifting out of frame? Usually because placement was never stated. Name the third, name the facing direction, and if the camera moves, describe where the subject sits at the start and at the end of the move.

How do I stop the model from inventing extra light sources? Describe the environment's light explicitly and close the door: "single source, no other lights visible." Saying what is absent is often as effective as saying what is present.

Should I use technical terms or plain language? Both. Terms like "85mm" or "rim light" anchor the model to real photographic references, while plain language like "cold light from a doorway behind her" supplies intent. Together they beat either alone.

How many generations should a good shot take? Two to four passes with one variable changed each time. Past eight attempts the card is underspecified — rewrite it instead of re-rolling.

Can composition be fixed after generation? Crops and reframes rescue minor issues, and a grade unifies colour, but you cannot add a foreground layer or reverse a light direction in post. Those belong in the prompt.

What if the scene needs a look I cannot name? Work backwards from a reference. Find an image with the framing, light, or colour you want, describe what you see in it, and paste that description into your card. Naming what you observe is easier than inventing vocabulary.

Give Your Next Shot a Point of View

Composition and lighting are not decoration applied after the fact. They are the sentence structure of the shot, the part that tells an audience what to feel before anything moves. Writing them down costs a few extra minutes per shot and saves entire afternoons of re-rolling.

Start with one scene. Write three shot cards with an explicit lens, subject position, key direction, and quality. Generate them in Orelon's AI video generator, pull framing and lighting references from the prompt library when a look is hard to put into words, and start from the video templates when you would rather adapt a proven structure than face a blank page. If you need a still to anchor a character or a palette, build it in the AI image generator first and carry it into every card.

Build the frame before the effect. Make each decision on purpose, and the model will follow.