Orelon logoOrelon
Preise

AI Text to Image and Video Prompts: A Practical Guide

7. Nov. 2026 · Von Orelon Team

KI-Video-Vorlagen entdecken

Lass dich von ein paar Community-Kreationen inspirieren und öffne dann eine Vorlage, um in Orelon weiterzuerschaffen.

Learn how to turn a written prompt into cinematic images and video: structure, consistency, camera language, iteration loops, and common mistakes.

A prompt is not a caption. It is a set of production decisions compressed into a paragraph. When someone types "a woman in a rainy city at night" and gets something flat back, the model rarely failed — the instruction simply left the cinematographer's work undone. No lens, no light source, no reason for the camera to be where it was.

The gap between a mediocre generation and a usable one is almost never the model. It is the specificity of the brief. This guide covers how to structure prompts for stills and motion, how to keep a character or location consistent across a sequence, how to iterate without destroying what already worked, and where the usual failure points hide. The techniques apply whether the frame is finished in an AI image generator or handed straight to an AI video generator.

Treat the prompt as a shot list, not a caption

A caption describes what is in the picture. A shot list describes how the picture gets made. That distinction changes everything about how you write.

  • "A detective in an office."
  • "A tired detective in his fifties at a cluttered desk, lit by a single green banker's lamp from the left, venetian blind shadows striping the wall behind him, 40mm lens at chest height, shallow depth of field, dust in the air, muted teal-and-amber grade."

Both describe the same idea. Only one tells the model where to stand, what to light, and what to leave soft. The second prompt is not longer because it is padded — it is longer because it contains decisions. Every decision you leave out is one the model makes for you, and models default to the middle of the distribution: even light, centered framing, generic texture.

The practical rule: if you would not accept a crew member choosing it at random, write it down.

The five layers of a prompt that holds up

Reliable prompts tend to contain the same five kinds of information. The order matters less than the presence of all five.

Subject and intent

Who or what is on screen, doing what, and why does the frame exist? "A cyclist" is a subject. "A cyclist pushing a loaded bike up a hill at dusk, exhausted but determined" is a subject with intent, and intent is what produces readable expression and body language.

Be concrete about age, wardrobe, and physical state when they affect the read of the shot. "A woman in her thirties in a damp wool coat" gives the model three separate anchors. "A person" gives it nothing.

Environment and light

Light is the highest-leverage element in any prompt. Name the source, the direction, and the quality: window, practical lamp, neon sign, overcast sky, fire, phone screen; backlit, side-lit from camera left, top-down; hard and directional, soft and wraparound, dappled, flickering.

Then place the subject inside that light. "Backlit by a low sun through tall grass, rim light on her hair, face in soft bounce" is a shot. "Sunset" is a time of day.

Camera and lens

Camera language is the fastest way to make a frame look intentional. Useful vocabulary: 24mm for wide and immersive, 50mm for neutral, 85mm for portraits with compressed backgrounds, 135mm for isolation. Height options run from ground level through eye level to overhead. Movement goes from static tripod to slow push in, handheld drift, orbit, crane up. Depth is either shallow with one sharp plane or deep with everything readable.

You do not need all of it in every prompt. Pick the one or two choices that carry the mood.

Style, texture, and grade

Style words are where prompts most often go wrong, because vague ones pull toward cliché. "Cinematic" alone means nothing — it has described everything from handheld documentary to anamorphic science fiction.

Describe what the image physically looks like instead: film grain, halation around highlights, soft lens flare, matte contrast, crushed blacks, crisp digital clarity. Reference a medium rather than a mood — 16mm film, large-format photography, watercolor wash, risograph print. Mediums carry texture; moods do not.

Motion and duration, for video only

When a prompt goes to a video model, add three more decisions: what moves, how much, and for how long. What moves can be the subject, the camera, or the environment — rain, smoke, cloth, hair. How much ranges from a subtle drift to a full dolly; restraint reads as confidence, while big movement on a short clip often reads as chaos. Duration should be planned around four-to-eight-second beats, with one continuous action per clip rather than a mini-narrative.

If nothing in the prompt implies motion, the model invents some, and it is usually the wrong kind. Say what stays still.

Still first, motion second

The most reliable AI video workflow is not to prompt video directly. It is to solve the frame as an image, then animate it. Most professional-looking AI sequences are built this way.

Pass one: lock the frame

Generate stills until composition, lighting, and character read are right. Stills are cheap to iterate and easy to compare side by side. Keep going until you have a frame you would publish on its own — then save the exact prompt that produced it. That prompt is now a reference document, not a one-off.

Pass two: add movement with restraint

Take the approved still in as the first frame and write a motion prompt that describes only the change: "slow push in, curtain drifts, dust motes rise, subject holds still." Do not re-describe the scene, because the frame already contains it. Re-describing invites the model to reinterpret the image rather than animate it.

Pass three: change one variable at a time

When a clip comes back wrong, the temptation is to rewrite everything. Resist it. Change one element — movement intensity, light direction, focal length — and regenerate. If you change four things at once, you learn nothing about which one mattered. This is slow at first and fast later, because you build a private vocabulary of what your phrasing actually does.

Keeping characters and sets consistent across shots

Consistency is the hardest problem in AI production and the one that separates a sequence from a pile of clips.

Anchor identity with a reusable description. Write one character block — age, build, hair, wardrobe, distinguishing feature — and paste it verbatim into every prompt that includes that character. Never paraphrase it. "Silver-streaked dark hair, scar through left eyebrow, olive field jacket" reproduces far better than three different phrasings of the same idea.

Anchor the world the same way. Location descriptions should be equally rigid: wall color, signage, time of day, weather, key props. A recurring location is a recurring paragraph.

Reuse seeds and reference images where the tool allows it. One approved frame does more for consistency than a hundred adjectives.

Shoot coverage, not variety. Generate the same scene from three angles at one focal length rather than three unrelated pretty frames. Coverage is what makes a cut feel like a scene.

Accept drift where it helps. Silhouettes, hands, over-the-shoulder frames, and inserts buy editing room without demanding facial consistency.

Copy-ready prompt patterns for five common jobs

These are starting points, not finished prompts. Substitute specifics.

Product hero shot. "[Product] centered on matte concrete, large softbox from camera right, hard rim light behind, faint reflection beneath, 85mm macro at f/8, deep focus, clean neutral grade, no props in frame."

Character portrait. "[Character block], seated, looking just off camera, single window from camera left, falloff into shadow, 85mm at eye level, shallow depth of field, fine grain, muted palette."

Establishing shot. "Wide view of [location], [time of day], [weather], layered foreground interest, 24mm at chest height, deep focus, atmospheric haze separating the planes, cool shadows with warm practical lights."

Motion insert. "Close-up of [object] on a table, static camera, steam rising slowly, shallow focus, soft directional light, six-second clip, no camera movement."

Transition beat. "Handheld shot following [subject] through a doorway into [second space], camera pushes forward at walking pace, light shifts from warm interior to cold exterior, natural motion blur."

Notice how much of each prompt is light and lens, and how little is adjectives. That ratio is the trick.

Where prompts fail: eight mistakes worth fixing

  1. Stacking style adjectives. "Cinematic, epic, stunning, 8K, masterpiece" adds noise, not information. Replace each with a physical property.
  2. Contradicting yourself. "Soft hard light" confuses the model into averaging both.
  3. Describing a story instead of a frame. A still cannot show a before and after. Pick the moment.
  4. Forgetting the background. An unmentioned background becomes generic blur. Say what is behind the subject.
  5. Overloading one prompt. Ten subjects means none render well. Split into shots.
  6. Ignoring the first frame. In video tools the source image does most of the work; a weak start cannot be rescued by a strong motion prompt.
  7. Changing everything at once. You lose the ability to attribute improvement.
  8. Not saving what worked. A documented prompt that produced an approved frame is the most valuable asset in the workflow.

Picking tools by the job, not by the leaderboard

Model rankings change monthly. Job requirements do not. Choose by these criteria instead: control over the first frame (if composition matters, prioritize reference-image support); motion predictability (test subtle camera moves against complex action with your own shot); iteration cost and speed (a slightly weaker tool you can run twenty times beats a stronger one you can only try twice); consistency features such as reference images, character locks, and seed control; resolution and aspect ratio options for vertical, wide, or square delivery; and export friendliness — codec, frame rate, and whether the clip grades well next to your other footage.

A useful exercise: run the same three prompts — a portrait, a wide establishing shot, and a motion insert — through any candidate tool. That fifteen-minute test tells you more than any comparison chart. If you are weighing specific platforms, the alternatives overview is a reasonable place to map the field.

A full production loop, end to end

Here is the loop in practice for a thirty-second sequence.

  1. Write the beat sheet. Six to eight beats, one sentence each. No visuals yet.
  2. Define the anchors. Character block, location block, palette, lens family. Write once, reuse everywhere.
  3. Generate stills per beat. Two or three variations on the same prompt skeleton.
  4. Select and lock. One frame per beat, with the exact prompt and settings noted.
  5. Animate each approved frame. One motion idea per clip, nothing more.
  6. Assemble and cut. Trim to rhythm, cut on motion rather than stillness.
  7. Patch. Regenerate only the shots that break the cut, using the saved prompts.

Most of the time goes into steps three and four, and that is correct. Solving the frame is the work; animating a solved frame is nearly mechanical.

FAQ

Do longer prompts always work better? No. Length is a side effect of specificity, not the goal. A forty-word prompt with a light direction, a lens, and a clear subject beats 120 words of mood adjectives. Cut any word that does not change the picture.

Should prompts be written differently for image and video models? In one specific way: video prompts need motion and duration, and they usually need less scene description when you supply a first frame. Repeating the full scene often causes the model to reinterpret the image instead of animating it.

How do I fix garbled hands, text, or faces? Reframe. Move the subject further from camera, crop the problem area, or change the angle. Prompt-level anatomy fixes are unreliable; composition-level fixes are not. For on-screen text, generate a clean plate and add typography in post.

Why does the same prompt give different results each run? Generation is stochastic. That variance is useful during exploration and painful during production, which is why saving the exact prompt, seed, and reference image matters. Reproducibility is a documentation problem more than a model problem.

How many variations before changing the prompt? Three or four. If none are close, the prompt has a structural problem, not a luck problem. If one is close, keep the prompt and vary a single detail.

Can one prompt work across multiple tools? The structure transfers; the vocabulary may not. Every engine responds more strongly to certain phrases. Keep a personal glossary per tool and expect to re-tune rather than copy-paste.

From prompt to finished cut

Everything here comes down to one habit: decide before you generate. Decide the light, the lens, the moment, the motion, and the reason the frame exists — then write those decisions down in plain language. Models are obedient to specificity and indifferent to enthusiasm.

When you are ready to practice, start with a single frame in the Orelon image generator, lock it until it holds up on its own, then carry it into the Orelon AI video generator and add one deliberate movement. For a head start on phrasing, the prompt library offers ready structures to adapt, and the templates show how finished sequences are assembled from individual beats. Cinematic ideas in motion start as a well-written sentence.