Orelon logoOrelon
料金

How to Write AI Image Prompts as Visual Blueprints That Work

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。

Build reusable AI image prompt blueprints: subject, lens, light, palette, and motion cues that carry cleanly from stills into cinematic video shots.

Most disappointing AI images are not a model problem. They are a description problem. You typed a sentence, the model filled in a thousand decisions you never made, and the result looked generic because it was generic — generic subject, generic light, generic lens. A visual blueprint fixes that by turning a wish into a specification.

The idea is simple: write prompts the way a director talks to a crew. Name the subject, the lens, the light, the palette, the surface texture, and the framing. Then reuse that structure until it becomes muscle memory. Once you can hold a look steady across several images, you can push the same language into motion and get shots that feel like they belong to the same film rather than the same afternoon.

Why a Blueprint Beats a Sentence

A single-sentence prompt works for quick exploration, but it collapses the moment you need consistency. If you ask for "a woman in a rainy city at night," you will get something. Ask for it again and you get a different something: different age, different city, different rain. Nothing is wrong with the model. You simply left every important choice open.

A blueprint closes those gaps deliberately. It also makes revision possible. When an image misses, a sentence-level prompt gives you nothing to fix except "try again." A blueprint gives you levers: the light was too flat, the lens was too wide, the palette drifted warm. You change one lever, regenerate, and compare. That is how professionals iterate — not by rewriting everything, but by adjusting a variable.

There is a second benefit that matters more over time. Blueprints are portable. The same structure works across different image models, and it survives the jump to video generation, where vague prompts fail even harder because motion adds another layer of ambiguity.

The Seven Slots Every Prompt Needs

Treat these as a checklist, not a rigid formula. Some slots will be one word; others will be a full clause. What matters is that you consciously fill or deliberately skip each one.

Slot What it controls Example
Subject Who or what carries the frame "a 40-year-old ferry captain in a wool coat"
Action or state What is happening right now "mid-turn, looking over her shoulder"
Environment Place, weather, era, density "pre-dawn harbor, fog sitting low on the water"
Light Source, direction, quality, ratio "single warm practical from the left, cold ambient fill"
Camera Focal length, height, angle, distance "85mm, chest height, medium close-up"
Palette and texture Color logic plus surface feel "desaturated teal shadows, salt-stained fabric, fine grain"
Finish Aspect ratio, medium, reference feel "2.39:1, photorealistic, subtle halation"

Write them in that order and your prompt reads like a shot note. Skip a slot on purpose when you want the model to surprise you — but know that you skipped it.

Casting the Subject Like a Casting Note

"A man" is not a subject; it is a category. Specificity is what makes an image feel observed instead of assembled. Age range, build, wardrobe texture, posture, and one small imperfection will do more for realism than any quality keyword.

Compare "a young businessman" with "a man in his early thirties, creased linen shirt with the top button undone, shoulders slightly forward, a healed cut on one knuckle." The second version tells the model what kind of story it is in. Small physical details also give the renderer something concrete to resolve, which reduces the plasticky smoothing that appears when a prompt is too abstract.

For characters you plan to reuse, write the subject block once and keep it verbatim across every prompt. Change only light, camera, and environment around it. This is the cheapest character consistency you can get, and it works far better than hoping the model remembers.

Directing the Camera

Camera language is the fastest way to move an image from snapshot to cinema. Focal length alone changes the emotional read: a 24mm frame feels immersive and slightly aggressive, a 50mm feels neutral and human, an 85mm compresses the background and pushes the viewer into the subject's space.

Then decide height and angle. Eye level is the default and therefore the least interesting. A camera at hip height makes a room feel tall. A slight top-down angle reads as vulnerability or surveillance, depending on the light. A low angle with a wide lens is a hero shot whether you intended it or not.

Finally, name the distance. "Wide establishing shot" and "medium close-up" produce visibly different compositions even with the same subject and light. If you want a specific crop, say it: "framed to the collarbone, subject on the right third, negative space on the left for titles." Framing for text is a professional habit worth building early, especially if the image will become a thumbnail or a video title card.

Lighting Is the Cheapest Upgrade

Lighting is where most prompts are laziest, and where the biggest gains hide. "Cinematic lighting" is a mood, not an instruction. Replace it with a source, a direction, a quality, and a contrast ratio.

  • Source: window, practical lamp, neon sign, overcast sky, firelight, phone screen.
  • Direction: side, back, top, under, three-quarter.
  • Quality: hard with crisp edges, or soft with gradual falloff.
  • Ratio: how dark the shadows go relative to the highlights.

"Lit by a single bare bulb above and slightly behind, hard shadows, deep falloff into the corners, cool spill from a doorway on the right" gives the model a lighting diagram. That sentence alone will outperform three paragraphs of adjectives.

A useful exercise: take one subject and write five prompts that differ only in lighting. Same person, same lens, same room. Watch how much of the emotional register is carried by light alone. It is usually more than you expect.

Color, Texture, and Finish

Color should be decided, not discovered. Pick a dominant, a secondary, and an accent, then constrain the saturation. "Warm amber key against cool blue-grey shadows" is a grade. "Vibrant" is not.

Texture is the second half of realism. Skin, fabric, metal, and glass all fail in different ways, and naming the material helps: brushed aluminum, wet asphalt with light pooling in the puddles, chipped paint over rust, matte ceramic with a thin glaze.

Finish is your final control. Aspect ratio, medium, and post-processing cues — 2.39:1 widescreen, 4:5 vertical, photorealistic, illustrated with visible ink hatching, 35mm film with fine grain and mild halation. Keep the finish block short and stable. If you change it every generation, you will never be able to tell whether the subject or the finish caused the difference you like.

Carrying the Blueprint Into Motion

Stills are the training ground. Video is where the discipline pays off, because motion multiplies ambiguity. A prompt that produced a beautiful portrait may produce a drifting, morphing mess when the model is asked to animate it without guidance.

Motion cues that survive the jump

Add a movement clause that names both camera and subject. "Slow dolly in, subject turns to camera and stops" is directable. "Dynamic movement" is not. Keep the number of simultaneous motions small — one camera move plus one subject action is usually the ceiling before output turns unstable.

Also name what should stay still. If the background is not meant to move, say so. If hair and fabric should react to wind, say that instead of assuming the model will infer it. And specify pacing: a slow push reads as tension, a quick handheld drift reads as documentary immediacy.

Continuity between stills and shots

When you move from a locked image to a sequence, keep the light and palette blocks identical and vary only camera and action. That single habit is the difference between a set of clips and a scene. Build a small reference sheet first — front, three-quarter, and profile of your subject, plus one wide of the location — then feed that language forward.

This is exactly the workflow Orelon is built around: develop the look as a still, then carry it into motion with the AI video generator while keeping the same descriptive blocks. If you want to develop the reference frame first, the AI image generator handles that stage.

A Repeatable Prompt Workflow

Draft in blocks. Write the seven slots on separate lines. Do not merge them yet.

Generate three variants. Change exactly one slot across the three. Compare, choose, and note what won.

Lock the winner. Copy the winning text into a notes file with a short label: "harbor captain, dawn, 85mm." That file becomes your personal library.

Test one variable at a time. New light, or new lens, not both. Otherwise you learn nothing.

Promote to motion last. Only animate a blueprint that already produces a still you would publish.

If you would rather start from proven structures than a blank box, the prompt library has reusable patterns you can adapt, and the video templates are useful when the deliverable is a finished cut rather than a single frame.

Three Blueprints, One Scene

Here is how the same blueprint flexes without losing identity.

Establishing: "Fishing harbor at pre-dawn, fog sitting low, a 40-year-old ferry captain in a wool coat walking the dock, 24mm, hip height, wide shot, single cold ambient sky light with warm practicals from boat cabins, desaturated teal shadows with amber accents, wet wood texture, 2.39:1, photorealistic."

Character: Same subject block. "85mm, chest height, medium close-up, slight left three-quarter, hard key from a cabin lamp, deep falloff, salt-stained wool visible, same palette, 2.39:1."

Motion: Same light and palette. "Slow dolly in from medium to close, captain turns from the water toward camera and holds, fog drifts right to left, no background movement beyond water, 2.39:1, subtle grain."

Three prompts, one filmic world. Nothing in the motion version contradicts the still, which is why the sequence holds together.

Mistakes That Flatten Results

  • Stacking contradictory cues. "Soft hard light" or "wide-angle portrait with compressed background" confuses the render.
  • Adjective piles. Five mood words do less than one lighting diagram.
  • Changing everything at once. You will find a better image and lose the reason.
  • Ignoring aspect ratio until the end. Cropping a hero composition to vertical often destroys it.
  • Overloading motion. Three simultaneous moves produce mush.
  • No reference sheet. Character drift across shots is almost always a documentation failure, not a model failure.

FAQ

How long should an image prompt be? Long enough to fill the seven slots, short enough to stay readable. Most strong prompts land between 45 and 90 words. Beyond that, you are usually repeating yourself rather than adding control.

Do quality keywords actually help? Rarely as much as structure. Words like "masterpiece" or "8K" add little once subject, lens, and light are specific. Concrete nouns beat praise.

Can I reuse one blueprint across different models? Yes, and you should. Expect small variations in how each model interprets lens and lighting language, then adjust one slot rather than rewriting the whole prompt.

How do I keep a character consistent across many images? Freeze the subject block word for word, keep the palette block identical, and vary only camera and action. Add a saved reference frame when the tool supports it.

What is the biggest difference between still and video prompting? Motion. Stills reward composition and light; video rewards explicit, limited movement plus a clear statement of what should stay still.

How many variations should I generate before deciding? Three is usually enough to reveal a direction. If all three miss, the blueprint is the problem — not the seed.

Build the Blueprint, Then Let It Move

A prompt is not a wish. It is a short, precise document that tells a system what to build. Write it in slots, lock what works, and change one variable at a time. Do that for a week and you will have something more valuable than any single good image: a repeatable way to describe the look you want.

When the still is right, take it into motion. Orelon is built for exactly that handoff — develop a cinematic frame, then animate it with the same language and keep the grade, the lens, and the mood intact. Start with a blueprint on the Orelon homepage, explore the Orelon blog for more workflow patterns, and see how the same descriptive discipline scales from one frame to a full sequence.