Orelon logoOrelon
料金

How to Prompt an AI Video Generator for Cinematic Results

2026年11月11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 a repeatable system for prompting AI video generators: shot anatomy, camera language, negative prompts, continuity, image-to-video, and fixes.

A prompt is not a wish. It is a shot brief — the same set of notes a director hands a cinematographer, a gaffer, and an editor before a take. When an AI video generator returns something that feels off, the model is rarely the whole problem; the brief was too thin to decide anything with. "Cinematic city at night" gives a system nothing to choose between a rain-slicked alley in Osaka and a helicopter pass over Los Angeles. This guide separates prompting into the parts you can actually control — subject, camera, light, motion, continuity, guardrails — and shows how to phrase each one so the model has something to obey.

Why Prompting Is Direction, Not Decoration

Every prompt contains two kinds of information: what the shot is about, and how it should be filmed. Most weak prompts only supply the first. They describe a thing and stop, leaving framing, lens, light, and movement to chance. The result is technically responsive but emotionally flat — a moving image with no point of view.

Directors solve this by giving notes. "Wide, locked off, late afternoon, let the shadows do the work" is a note. So is "handheld, close, breathing distance, keep focus on her hands." Notes are decisions, and decisions are what make footage feel authored rather than generated.

Treat prompting as note-giving and three practical things change. First, you stop rewriting entire prompts and start adjusting single clauses, which makes iteration fast. Second, you can reuse the same subject across ten variations by swapping only the camera line. Third, you build a personal vocabulary — phrases you know produce a specific look — and that vocabulary is what separates people who get consistent output from people who get lucky once.

The takeaway before you type: decide the shot. Where is the camera? What is the light doing? What moves? How long is this moment? Answer those four questions and the prompt largely writes itself.

The Anatomy of a Cinematic Prompt

A dependable prompt has five slots. Not every shot needs all five in full, but knowing the slots stops you from leaving something important blank.

Subject and action

Be concrete about who or what is on screen and what they are doing at this exact second. Nouns alone are weak; nouns plus a verb in progress are strong. "A cyclist" is thin. "A cyclist coasting downhill, hands off the bars, jacket snapping" gives the model a body, a direction, and a physical consequence.

If there are multiple subjects, state their relationship: "a father and daughter walking, she pulls ahead, he slows to let her." Relationship language produces staging. Lists of people produce a crowd with no story.

Camera and framing

This is the slot most people skip and the one that changes results fastest. Specify shot size (extreme wide, wide, medium, close, macro), angle (eye level, low, high, overhead, dutch), and stability (locked off, handheld, gimbal glide, drone). Add a lens feel when it matters: wide-angle distortion, long-lens compression, shallow depth of field.

Compare "a chef plating a dish" with "medium close-up, slight low angle, 50mm feel, shallow focus on the plate, hands sharp, background kitchen falling soft." The second sentence is a decision. The model can execute a decision.

Light, color, and texture

Light is where cinematic quality actually lives. Name the source and the quality: hard afternoon sun, overcast diffusion, practical neon, single window light, bounced bounce off a white wall. Then name the palette in plain words — warm amber and deep brown, cold cyan with a single red accent, desaturated greens.

Texture covers the finish: fine grain, clean digital, slight haze, wet surfaces, dust in the air. A useful rule is one light idea plus one palette idea. Stacking five colors produces mud.

Motion, pacing, and duration

Say what moves and how fast. "Slow push in," "gentle drift left," "subject walks toward camera, background static," "time-lapse of clouds over a fixed skyline." Motion verbs set tempo; tempo sets mood. A slow push reads as tension. A snap zoom reads as comedy or urgency.

Keep single generations short and purposeful. A four-to-eight second beat with one clear action almost always beats a long prompt trying to compress a whole scene into one clip.

Mood and audio intent

Mood words are the least precise slot, so use them sparingly and pair them with physical detail. "Melancholy" alone is vague; "melancholy — empty bus stop, wet pavement, one flickering fluorescent" gives the model something to render. If the tool supports audio or you plan to add sound later, note the intent: "quiet room tone," "distant traffic," "no music, let the footsteps carry it."

A Layer-by-Layer Workflow You Can Repeat

The four-pass draft

Write prompts in four passes instead of one. Pass one: subject and action only. Generate a still or short clip and check whether the idea reads. Pass two: add camera and framing. Pass three: add light, palette, and texture. Pass four: add motion and duration.

This is slower on paper and much faster in practice, because when something breaks you know which layer caused it. If the framing is wrong at pass two, you never waste time debugging it after adding six other variables.

A reusable shot block

Keep a personal template with empty slots and fill it per shot:

[shot size and angle], [subject] [action in progress], [setting detail], [light source and quality], [palette], [texture or finish], [camera movement], [duration feel]

Written out, a finished block looks like this:

Medium wide, low angle. A lone swimmer pushes off the wall in an empty indoor pool, water still settling. Late-night fluorescent overhead, greenish cast. Cool palette with one warm exit sign. Slight haze, wet tile reflections. Slow lateral drift right. Calm, unhurried.

Save variations of that block in a notes file or a prompt library so your best phrasing survives between projects.

Negative Prompts and Other Guardrails

Negative prompts are instructions about what must not appear. They are most useful when a model has a persistent habit that keeps ruining otherwise good shots — warped hands, text artifacts, extra limbs, watermarks, jittery edges, plastic skin, oversaturated skies.

Use them surgically. A short list of the three to five problems you actually see beats a giant block of prohibitions, which can flatten the image or cancel legitimate detail. If you forbid "blur," you may also kill the depth of field you wanted.

Beyond negatives, three guardrails do a lot of quiet work. Constrain the number of subjects ("one person, no background figures"). Constrain the setting ("interior only, no windows"). Constrain the era or technology level when the model keeps adding anachronisms. Each constraint removes a decision the model would otherwise make for you — usually badly.

Temporal Consistency: Keeping a Shot From Drifting

Drift is the classic failure of AI video: the face changes, the jacket color shifts, the background morphs into something else halfway through. You cannot eliminate it entirely, but you can reduce it substantially.

Continuity anchors

Anchor identity with two or three stable descriptors repeated in every prompt for that shot: a specific garment color, a hairstyle, a distinguishing object. "Grey wool coat, short dark hair, small scar above the left eyebrow" repeated across prompts keeps a character recognisable far better than "a woman" ever will.

Anchor the environment the same way. If the scene is a diner, name the same details each time — chrome counter, red vinyl stools, handwritten menu board — so the model has fixed reference points instead of improvising a new diner each generation.

Shot length and motion verbs

Short shots drift less than long ones. If a clip must run longer, keep the camera and subject doing one continuous thing rather than introducing new action mid-shot. A slow push on a static subject is stable; a push that becomes a pan that becomes a character turn is three chances to break.

Choose motion verbs that imply continuity: drift, glide, settle, follow, hold. Verbs like "spin," "morph," or "transform" invite the model to rebuild the frame, which is exactly where consistency dies.

Reference frames

When a shot absolutely must match a previous one, generate from a reference image instead of text alone. A single strong frame is a far tighter constraint than a paragraph, and it removes most ambiguity about style, wardrobe, and lighting.

Image-to-Video and Reference Frames

Image-to-video is the highest-leverage technique in the whole workflow, and it is underused because people treat it as a shortcut rather than a control.

The pattern: generate or photograph a still you love with an AI image generator, then use it as the first frame and let the prompt describe only the movement. Instead of re-describing the scene, write "slow push in, dust drifting through the light, curtain moving slightly." The model inherits composition, palette, and detail from the frame and spends its effort on motion, which is where you want it.

Three practical notes. Compose stills with motion in mind — leave space where the camera will travel, avoid cropping limbs that will need to move. Keep the prompt's style language consistent with the reference, or the model will fight itself. And for multi-image workflows, use one image for subject identity and a second for look or environment, describing their blend explicitly: "character from the first image, lighting and palette from the second."

If you are producing a sequence rather than a single shot, this is also how you keep a series coherent: one reference frame per setup, reused across every generation of that setup.

Prompt Patterns by Use Case

Different formats reward different prompt habits. Four patterns cover most work.

Product and brand film

Prioritise precision over drama. Name the surface (brushed aluminium, matte ceramic), the light (softbox from the left, hard rim from behind), and the motion (slow orbit, rack focus from logo to detail). Keep the background controlled and simple. Product shots fail when the model invents an environment; say "seamless grey backdrop" and mean it.

Character and dialogue beats

Prioritise face, eyeline, and micro-movement. Shot size medium close or close. Specify where the subject is looking and what their hands are doing, because hands and eyeline carry most of the performance. Keep the action small — a look, a breath, a half-smile — and let the camera hold.

Landscape and establishing shots

Prioritise atmosphere and scale. Wide or extreme wide, slow movement, weather named explicitly. Give the model a foreground element to anchor depth: a rock edge, a fence post, a doorway. Without a foreground, wide shots read as flat backdrops.

Abstract and brand loops

Prioritise rhythm and palette over realism. Describe the material behaviour (liquid metal folding, ink dispersing, light ribbons twisting) and the loop logic ("continuous, seamless return to start"). Abstract work tolerates more ambiguity, so this is where you can experiment with unusual colour pairings and let the model surprise you.

Common Mistakes and How to Fix Them

Symptom Likely cause Fix
Output feels generic No camera or light decisions Add shot size, angle, and one light source
Subject changes mid-clip Too many actions in one prompt Shorten the shot, use one motion verb
Image looks plastic No texture language Add grain, haze, or surface detail
Colours look muddy Too many competing palettes Limit to one dominant colour idea
Hands and text break Model limitation Reframe the shot or use negative prompts
Every generation is different No reference frame Anchor with a still image
Prompt stops working when lengthened Clause conflicts Split into two shots instead of one long line

A last, less obvious mistake: chasing a single perfect prompt instead of a reliable prompt family. Five near-identical variations with one variable changed will teach you more in ten minutes than an hour of rewriting a paragraph from scratch.

FAQ

How long should an AI video prompt be? Long enough to cover the five slots, short enough that no clause contradicts another. For most shots that lands between 25 and 60 words. If you pass 80 words, look for redundant mood adjectives to cut.

Should I put the camera first or the subject first? Either works, but be consistent within a project. Camera-first phrasing tends to produce more controlled composition; subject-first phrasing tends to produce more energetic action. Test both on one shot and keep the one that matches your intent.

Do negative prompts actually help? Yes, when they target a repeated, specific failure. They hurt when they become a long list of fears. Start with nothing, note what breaks across five generations, then add only those items.

How do I keep a character consistent across shots? Three things together: a fixed written description, a reference frame used as the first image, and a consistent light and palette description. Text alone drifts; text plus reference holds.

Why does my prompt work sometimes and fail other times? Usually because a clause is ambiguous and the model resolves it differently each run. Rewrite the ambiguous phrase into a physical description — "warm light" becomes "low sun through a window, amber cast."

Can I reuse one prompt for a whole sequence? Reuse the style block and change only the action and camera lines. That gives you a coherent look with distinct beats, which is exactly how a real sequence is built.

Put Your Prompt to Work in Orelon

You now have a system: five slots, four passes, short shots, continuity anchors, and reference frames doing the heavy lifting. The fastest way to make it yours is to run it on real footage. Open the AI video generator and build one shot in four passes rather than one long paragraph. Then generate three variations with a single clause changed and compare them side by side. Browse video templates when you want a starting structure, or read more workflow breakdowns on the Orelon blog between sessions. Cinematic ideas in motion reward direction — give the model something specific to obey, and it will give you something worth watching.