Orelon logoOrelon
料金

Prompt to Video: How to Turn Ideas Into Cinematic Scenes

2026年11月5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 to write prompts that produce cinematic AI video: shot structure, camera language, lighting, continuity, iteration loops, and fixes for common failures.

You have the shot in your head: the low sun behind the tree line, a figure walking away, dust catching the light. What you do not have is a camera crew, a location, or three weeks of post-production. Text-to-video closes most of that gap, but only if you stop treating the prompt as a magic sentence and start treating it as a shot list compressed into a paragraph.

That is the whole craft. A prompt is not a wish. It is a set of decisions about subject, framing, movement, light, and duration, written in the order a director would call them out on set. Get that order right and the output stops feeling like a slot machine.

This guide covers what a prompt actually controls, a reusable skeleton you can adapt to any project, how to build sequences instead of isolated lucky clips, and how to diagnose the five failures that eat most of a beginner's time.

Start With a Shot, Not a Script

The most common mistake is writing a story. A paragraph about a lonely astronaut who misses Earth is a concept, not a shot, and the model has no way to choose which frame of that story you wanted.

Before you type anything, answer three questions in plain language:

  • What is the single most important thing in frame? One subject, one action.
  • Where is the camera? Distance, height, angle, and what it is doing.
  • What is the light doing? Time of day, direction, hardness, color.

If you cannot answer those three, the prompt will not either. Write them down as a sentence before you open any tool. This takes ninety seconds and saves dozens of regenerations.

A useful test: read your prompt aloud and ask whether a camera operator could set up the shot from it. If they would have to ask a follow-up question, the model will guess — and guessing is where mush comes from.

What a Prompt Actually Controls

Text-to-video systems map language onto visual probability. They respond strongly to nouns and physical descriptors, moderately to adjectives, and weakly to abstractions like 'emotional' or 'epic'. If you want a feeling, describe the physical evidence of that feeling: wet pavement, a clenched hand, fluorescent light on a tired face.

The four layers of a video prompt

Every effective prompt contains four layers, usually in this order:

  1. Subject and action. Who or what, doing exactly what, in one clause.
  2. Environment. Location, weather, time of day, background elements that matter.
  3. Camera. Shot size, angle, lens feel, and movement.
  4. Style and rendering. Film stock, grain, palette, reference era, level of realism.

When output disappoints, identify which layer failed rather than rewriting everything. If the framing is wrong, change only the camera sentence. If the mood is wrong, change only the light. Rewriting all four layers at once teaches you nothing about what worked.

Attention is finite

Every model has a practical attention budget. A prompt with twelve subjects will produce a blurry average of all twelve. Two or three visual ideas per shot is the sweet spot. Save the rest for the next shot — that is what a sequence is for.

A Reusable Prompt Skeleton

Here is a skeleton that works across subjects. Fill the brackets, keep the order, and delete anything you do not need.

[Shot size] of [subject] [action], in [environment] at [time of day],
[lighting description], [camera movement], [lens or film character],
[color palette], [duration or pacing note]

A filled example:

Slow dolly-in, medium close-up of a baker pulling a tray from an oven,
in a narrow tiled kitchen just before dawn, warm light spilling from the
oven door and cool blue window light from the left, shallow depth of field,
35mm film grain, muted amber and slate palette, unhurried pacing

Notice what is missing: no adjectives about how beautiful it is, no backstory, no camera brand names. Every word changes pixels.

Keep a prompt file

Save every prompt that produces a usable result, plus a one-line note on what you changed. After ten projects you will have a personal library that is worth more than any generic list, because it encodes your taste. Browsing a prompt library is a good starting point, but the notes you write yourself are what compound.

Camera and Light: Directing With Words

Camera language is the highest-leverage vocabulary in text-to-video, because it changes composition rather than decoration.

Shot size

  • Extreme wide establishes scale and isolation. Good openers.
  • Wide gives spatial context and lets action breathe.
  • Medium is the workhorse for people doing things.
  • Close-up carries emotion and detail.
  • Macro is for texture and product truth.

Camera movement

Movement is where generations most often break. Simple, physically plausible moves survive: slow push in, slow pull out, a gentle lateral track, a static frame with subject motion. Complex choreography — orbit while craning while racking focus — usually produces warping. If you need a complicated move, split it into two shots and cut between them.

Lighting as a directing tool

Describe light the way a gaffer would: source, direction, quality, color.

  • Source: window, practical lamp, neon sign, fire, overcast sky.
  • Direction: backlit, side-lit, top-lit, frontal.
  • Quality: hard and defined, soft and wrapped, dappled.
  • Color: tungsten warmth, daylight neutrality, sodium orange, moonlit blue.

'Golden hour backlight with lens flare' and 'cool overcast side light' produce completely different films from the same subject. If you only improve one thing in your prompts this month, improve the lighting sentence.

Building a Sequence, Not a Clip

A single clip is a demo. A sequence is a video. The difference is continuity, and continuity in AI video is a design problem, not a luck problem.

Continuity tactics that actually work

Write a character sheet. Before generating, write two sentences describing your subject that never change: age range, build, hair, wardrobe, distinguishing feature. Paste those sentences verbatim into every prompt. Consistency comes from repetition, not from memory.

Lock the look. Decide on palette and grain once, then repeat the style clause in every shot. Mixed styles are the fastest way to make a sequence feel assembled rather than directed.

Generate a still first. For anything with a recurring subject or product, produce a reference image, then use it as the visual anchor for video generations. Starting from an approved frame is far cheaper than discovering the design in motion. An AI image generator is useful here as a pre-production sketch tool.

Cover your cuts. Generate a few extra frames of environment, hands, textures, and empty rooms. These cutaways hide transitions and buy you rhythm in the edit.

Respect screen direction. If your subject exits frame right, the next shot should feel like it continues right. Breaking direction reads as a mistake even when viewers cannot name it.

The Iteration Loop: Draft to Final Take

Professional results come from a loop, not a single call. A loop that works:

  1. Draft wide. Generate four to six variations of the same prompt at low effort. You are looking for one that is structurally correct, not beautiful.
  2. Diagnose. Name the single biggest problem: subject wrong, framing wrong, motion broken, style off.
  3. Change one variable. Fix only that problem. Keep everything else identical so you learn the model's response.
  4. Refine the survivor. Once structure is right, spend your remaining attempts on polish: light quality, texture, micro-movement.
  5. Log it. Write the final prompt down with the winning settings.

Two rules keep this efficient. Never polish a shot whose framing is wrong — you will throw it away. And never accept a shot with a broken motion artifact because the lighting is pretty; artifacts are what audiences notice.

Failure Modes and Fast Fixes

Morphing faces and hands. Usually caused by too much motion or too small a subject in frame. Fix by moving the camera closer, reducing movement to a slow push, or adding a clear foreground element that anchors the frame.

Warping architecture and text. Straight lines and lettering are hard. Avoid signage as a focal point, shoot angles that hide long straight edges, and keep the subject closer to camera than the problem geometry.

Mushy, averaged frames. Almost always a prompt overload. Cut the number of subjects, remove contradictory style words, and drop adjectives that do not describe anything visible.

Style drift across shots. You changed the style clause between generations. Freeze it into a copy-paste block.

Dead, static feeling. You described a scene but not an action. Give the subject something small and specific to do: lift a cup, turn a page, step over a puddle, exhale.

Uncanny smoothness. Add imperfection language: handheld sway, slight grain, natural skin texture, imperfect focus falloff. Realism lives in the flaws.

Worked Example: A 30-Second Product Teaser

Suppose you are promoting a ceramic coffee mug. Six shots, roughly five seconds each.

  • Shot 1 — establish. Wide, macro-to-wide pull out from condensation on the ceramic surface to a wooden table, morning window light from the right.
  • Shot 2 — texture. Macro, slow lateral track across the glaze, hard raking light revealing texture, shallow depth of field.
  • Shot 3 — action. Medium close-up, hands lifting the mug, steam rising, backlit so the steam glows.
  • Shot 4 — context. Wide, the mug on a desk beside a notebook, cool daylight, subject small in frame for breathing room.
  • Shot 5 — detail. Close-up, a thumb resting on the rim, warm practical lamp light, soft falloff.
  • Shot 6 — close. Slow push in on the mug alone, lights dimming, single warm highlight.

Every shot shares the same palette clause and the same grain clause. The result cuts together because the design was decided before generation, not after. If you want a faster start, pre-built video templates can supply the structure so you only write the subject and light layers.

Choosing a Workflow That Fits Your Project

Not every project deserves the same process.

Social clips and thumb-stopping shorts. Speed matters most. One strong subject, one camera move, bright light, generate wide, pick the best, ship. Do not build a character sheet for a six-second clip.

Narrative scenes. Continuity matters most. Reference images, character sheets, locked style blocks, extra cutaways, and a genuine edit pass.

Product and brand work. Precision matters most. Macro detail shots, controlled lighting language, and a strict palette. Expect to regenerate more and to reject more.

Concept and pitch visuals. Clarity matters most. Favor wide, legible shots that communicate an idea in three seconds over atmospheric shots that need explanation.

If you are weighing tools rather than technique, read comparisons rather than feature lists — for example a breakdown of Orelon vs Runway or Orelon vs Kling AI — and test the same three prompts in each before deciding. The prompt is the constant; the model is the variable.

FAQ

How long should a prompt be?

Long enough to cover the four layers, short enough to stay coherent. For most shots that is 25 to 60 words. If yours runs past 90 words, you are probably describing two shots.

Do camera brand names or film stock names help?

Film stock language often helps because it implies grain, contrast, and palette. Camera brand names help less and sometimes confuse the model. Describe the look, not the gear.

Why does my second shot not match my first?

Because you described it from memory instead of copying the subject and style clauses verbatim. Copy-paste those two blocks every time.

Can I fix a clip with editing instead of regenerating?

Sometimes. Trimming the first and last half-second removes most transition artifacts, and cutaways hide broken motion. But if the subject's identity is wrong, regenerate — no edit fixes that.

Is a reference image better than a long prompt?

For anything consistent or branded, yes. A reference image communicates design in one glance what a paragraph struggles to specify. Use prompts for motion and light, images for identity.

How many attempts should a good shot take?

For simple subjects, three to six. For complex action or tricky geometry, expect ten or more. If a shot is consistently failing after fifteen attempts, the concept is fighting the medium — redesign the shot.

Bring Your Idea to Orelon

A prompt is a decision written down. The directors who get reliable results are not writing better poetry — they are making clearer decisions about subject, camera, light, and duration, then iterating on one variable at a time.

Start with a single shot you can describe in one sentence. Write the four layers, generate a draft, name the biggest problem, and change one thing. When you are ready to put that loop to work, Orelon's AI video generator is built for cinematic ideas in motion — and when you want to see how other creators structure their shots, the Orelon blog is a good next stop.