Orelon logoOrelon
Tarifs

Prompt Engineering for AI Visuals: A Creator's Workflow Guide

6 nov. 2026 · Par Orelon Team

Explorez les modèles vidéo IA

Parcourez quelques créations de la communauté pour trouver l’inspiration, puis ouvrez n’importe quel modèle pour continuer à créer dans Orelon.

Learn how to write prompts that produce consistent, cinematic AI images and video shots, with a repeatable workflow from first idea to final cut.

A prompt is not a magic phrase. It is a short specification, and like any specification it only works when the person writing it knows what the system on the other side can actually hear. Most disappointing AI images are not the model's fault; they are the fault of a request that described a feeling and expected the machine to infer a photograph. This guide walks through a practical, repeatable way to write prompts for still images, then carry those same ideas into motion so that your stills and your shots belong to the same visual world instead of looking like two unrelated projects.

The approach here is deliberately unglamorous. You will build a small visual vocabulary, test one variable at a time, and keep notes. That is what separates a creator who gets usable frames on the third attempt from one who burns an afternoon re-rolling the same vague sentence.

What a Prompt Actually Controls

Text-to-image systems encode your words into a numeric representation, then steer a sampling process toward regions of the model's learned visual space that correspond to that representation. Diffusion models are a good reference point if you want the mechanics; the practical takeaway matters more than the math: concrete, observable language moves the output reliably, while abstract intent moves it barely at all.

That single fact explains most prompt frustration. Words like "premium," "engaging," or "modern" carry almost no visual signal on their own. "Brushed aluminium body on a bone-white surface, single hard light from the upper right, faint contact shadow" carries a great deal. You are not describing what you want the viewer to feel; you are describing what the camera would see if the scene existed.

Think of a prompt as having six dials: subject, composition, light, palette, finish, and mood. In practice, the first four do about eighty percent of the work. Mood is usually a by-product of light and palette, and finish is the salt you add at the end rather than the base of the dish.

The Anatomy of a Strong Visual Prompt

Subject and action

Name who or what is in frame, what they are doing, how many of them there are, and what is immediately around them. "A cyclist" is a category. "A courier in a rain-soaked shell jacket, mid-pedal, water spraying off the rear tyre, a blurred delivery box strapped to the frame" is an image. Counts and spatial relationships are worth stating explicitly, because models quietly invent extra people, extra fingers, and extra chairs when you leave the inventory to chance.

Light and atmosphere

Lighting is the highest-leverage clause in almost any prompt. Compare "a kitchen" with "a kitchen at low winter sun, light raking across the counter and throwing a long window frame shadow onto the tiled floor." The second describes a moment. Good lighting phrases are physical: the source, its direction, its softness, its colour. "Single softbox forty-five degrees camera left," "overcast blue hour with wet asphalt reflecting shop signs," "late afternoon sun through blinds, hard stripes across the wall."

Camera, lens, and framing

Phrases like "85mm portrait, shallow depth of field," "24mm wide angle from a low camera position," or "overhead flat lay" nudge perspective, compression, and blur. They are not a literal camera simulation, so do not expect perfect optical physics. What they reliably do is push the composition toward a recognisable photographic idiom instead of a flat illustration.

Finish, texture, and grade

This is where restraint pays. "Fine film grain, matte highlights, muted teal shadows" is a clear instruction. Stacking twelve style adjectives produces mud, because the clauses start contradicting one another. Pick a finish and commit: clean commercial render, documentary 35mm, soft watercolour wash, high-contrast monochrome.

Negative prompts and what to leave out

Negative prompts are best used for artifacts, not for taste. "Extra fingers, warped hands, watermark, embedded text, duplicated limbs" is a sensible list. Long negative essays tend to fight the positive description and flatten the image. If you find yourself writing a paragraph of things you do not want, the positive prompt is usually under-specified.

A Worked Example: From Vague to Cinematic

Start with the kind of sentence most people type first:

a woman in a forest

It will produce something competent and forgettable. Notice what is missing: time of day, weather, lens, palette, action, wardrobe, and any reason for the frame to exist.

Second pass adds subject detail and light:

a woman in her thirties walking through a pine forest, early morning mist, soft diffused light

Better, still generic. The light is doing work now, but the composition is undirected.

Third pass adds camera and palette:

a woman in her thirties in a worn olive raincoat walking away from camera through a dense pine forest, early morning mist, soft diffused light, muted green and grey palette, 35mm lens, medium shot from behind at eye level

Now the frame has a point of view. The final pass adds finish and a specific moment:

Medium shot from behind at eye level, 35mm lens: a woman in her
thirties in a worn olive raincoat walking away from camera along
a narrow pine-forest trail, boots sinking slightly into wet
needle litter, early morning mist between the trunks, soft
diffused light from the left, muted green and grey palette with
one warm rust accent on her scarf, faint film grain, matte
highlights, shallow depth of field, no visible text

Read that aloud and you can picture the shot. That is the test. If you cannot describe the frame to a cinematographer, the model will not guess it for you.

Building Consistency Across a Shot List

A single striking image is a demo. A set of images that look like they came from the same production is a deliverable. Consistency comes from reuse, not from luck.

Character consistency

Write one descriptive block for each recurring character — age range, build, hair, wardrobe, distinguishing detail — and paste it verbatim into every prompt. Paraphrasing between prompts is the fastest way to make the same character look like a different actor in shot three. Where the tool supports seeds or reference images, reuse them alongside the text.

Environment and palette

Define the palette once, in words, and keep it. "Dusty terracotta, olive, bone white, one saturated accent" is a rule you can apply to a living room, a market street, and a close-up of hands. Without a stated palette, each prompt drifts toward whatever the model considers default, and the set starts to look assembled from different films.

Props, signage, and wardrobe

Physical continuity details are easy to forget and easy to fix. If a character carries a canvas bag in the wide shot, say so in the close-up. If the scene is set in a specific season, state the season in every prompt rather than assuming the model remembers.

From Stills to Moving Shots

Translating a still prompt into a video prompt is mostly a matter of subtraction plus one addition. You subtract the overloaded style clauses, because motion generation already introduces complexity, and you add a single, readable movement.

Good motion prompts describe one action and one camera behaviour within a short duration. "She turns her head toward the window, camera holds static, mist drifting through the frame" is achievable in a few seconds. "She walks through the forest, then cuts to a market, then looks at the camera" is three shots wearing a trench coat, and none of them will land cleanly.

Practical rules that hold up across most current video tools:

  • One movement per shot. Subject movement or camera movement, rarely both at full strength.
  • Name the camera behaviour explicitly. Locked-off, slow push in, gentle handheld drift, orbit left. Silence on this point usually produces an unplanned drift.
  • Keep duration honest. A five-second clip cannot contain a full narrative beat. Write for the length you have.
  • Reuse the still prompt as the opening frame description. If you generated a hero image first, describe that exact frame so the motion starts where the still left off.

When you are ready to test this in practice, the AI video generator is the natural place to convert a locked still prompt into a moving shot, while the AI image generator handles the frame-building stage before that.

A Repeatable Workflow in Seven Steps

  1. Write the brief as one sentence. Who is in frame, where, and what the audience should notice. If you cannot do this, the prompt cannot rescue the idea.
  2. Build a visual bible block. Subject description, palette, lighting philosophy, preferred lens, finish. This block gets pasted into every prompt in the project.
  3. Generate wide, cheap variations first. Six to eight rough frames at low effort, judged only on composition and light. Do not polish a bad composition.
  4. Change one variable at a time. Once you have a candidate, adjust only lens, or only light direction, or only palette. Multi-variable edits teach you nothing about what worked.
  5. Lock the look and record it. Save the exact prompt text, seed, aspect ratio, and settings next to the winning frame. Screenshots are not a record; the text is.
  6. Convert to motion one shot at a time. Add a single movement, keep duration tight, and check whether continuity with the previous shot survives.
  7. Assemble before regenerating. Cut the sequence together first. Many shots that look weak in isolation read perfectly in context, and regenerating them wastes time.

A prompt library speeds up step two considerably, because you are adapting a proven structure rather than starting from a blank field.

Common Mistakes That Wreck Otherwise Good Prompts

Adjective stacking. Five style words blur into an average of five aesthetics. Two is usually plenty.

Internal contradiction. "Wide-angle close-up," "soft hard light," and "minimalist maximalist" all push the sampler in two directions and land in the middle.

Missing counts and relationships. If it matters that there are two chairs and one is empty, say it. Otherwise you will get three chairs and a vase.

Treating the prompt like a search query. Keywords without syntax produce a collage of unrelated elements. Sentences with spatial logic produce compositions.

Changing everything at once. You get a better image and no idea why, which means you cannot repeat it.

Ignoring aspect ratio. A prompt written for a vertical social frame will crop badly in a 16:9 sequence. Choose the frame shape before you refine the content.

Expecting reliable text rendering. Signage and legible typography remain fragile. Design around them or add type in post.

Output Settings That Change the Look

Prompt text is only part of the equation. Aspect ratio determines what the composition can even contain — vertical favours subjects, horizontal favours environment. Resolution affects how much fine detail survives, which matters if you plan to push in during an edit. Guidance or strength settings trade prompt obedience against natural variation; high obedience gives literal, sometimes stiff results, lower values give more photographic looseness and less control. Seeds let you hold a composition steady while you iterate on wording.

Spend a few minutes testing these dials on one prompt before you write a hundred prompts. Knowing how your tool behaves at the extremes makes every later decision faster.

Keep a Prompt Log

A simple table beats memory. Columns: idea, prompt text, settings, verdict, next change. Two weeks of logging turns guesswork into a personal style guide, and it makes handoffs to editors or collaborators far easier. When a client asks for "the same look but for a winter campaign," you will have the exact language that produced the look.

Reusable video templates can also serve as structural starting points, especially for formats you repeat often, such as product rotations, title sequences, or talking-head inserts.

FAQ

How long should a prompt be?

Long enough to specify subject, light, composition, and finish — usually two to four sentences. Beyond that, you are often repeating yourself or contradicting earlier clauses. Length is not a virtue; specificity is.

Do style keywords like "cinematic" actually do anything?

They carry weak signal on their own. "Cinematic" is a mood word. "Anamorphic 40mm, shallow focus, warm practical lights in the background, crushed but not clipped shadows" is the same intention expressed as observable properties, and it works far more reliably.

Should I generate images before video, or go straight to motion?

For anything narrative, build the frame first. Locking a composition as a still is cheaper and faster to iterate, and it gives you a reference to describe when you start the motion pass.

Why does my character change between shots?

Almost always because the description changed. Paste the identical character block into every prompt, reuse seeds or reference images, and keep wardrobe details explicit rather than implied.

How many variations should I generate before judging?

Six to eight at low effort is a reasonable first pass. Judge composition and light only. Refine the best one rather than continuing to roll the dice on the original.

Can I reuse one prompt across different tools?

Broadly yes, but expect differences in how each interprets lens language and negative prompts. Treat prompts as portable drafts, not files that transfer perfectly.

Turn Your Prompt Into a Moving Image

The gap between a decent AI image and a coherent piece of video is not secret syntax. It is a repeatable process: describe what the camera sees, reuse your descriptive blocks, change one variable at a time, and record what worked so the next shot inherits it.

Once your still prompt is locked, take it into Orelon and give it motion. Start with a single shot, one movement, honest duration, and see how much of the frame's atmosphere survives the transition. Then build the sequence shot by shot — and browse the Orelon blog for more workflow breakdowns when you want to push the look further.