Orelon logoOrelon
Tarifs

AI Video Prompts: Direct Subject, Camera, Light, and Motion

11 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 AI video prompts that control subject, camera, light, and pace, with a reusable skeleton, a continuity kit, and a debugging checklist.

A prompt box is not a search bar. It is a shot list handed to a crew that follows instructions exactly and improvises everything you left out. That is why two people typing into the same generator can end up with footage that looks a decade apart: one wrote a wish, the other wrote a shot.

The encouraging part is that prompt writing for video is a learnable craft with a short ramp and a long refinement tail. What follows is a working system rather than a list of magic phrases: six controllable dials, a reusable skeleton, a continuity kit for anything longer than a single clip, a symptom-first debugging table, and the decision rules that tell you when to re-roll a clip instead of rewriting it. The same approach holds up for a three-second social loop and a twelve-shot brand film.

The six dials that shape every render

A generator is solving several constraints at once: what exists in frame, how it changes, where the camera sits, what the light is doing, and how long the whole thing holds together before its memory of the scene starts to drift. When those constraints fight each other, you get either a stiff tableau or a melting collage. Six dials cover the useful territory.

Subject: cast a person, not a category

"A woman in her thirties" is a casting note. "A woman in her thirties with cropped dark hair, a charcoal wool coat, and a scuffed leather satchel" is a picture. The difference matters because the model must choose a face, a silhouette, and a texture from whatever you handed it. Give it three visual specifics and it stops averaging.

The same rule applies to objects. "A car" becomes "a boxy hatchback with a rusted rear wheel arch." "Food" becomes "a bowl of ramen with a soft-boiled egg cut in half." Specificity is not decoration; it is the line between a stock asset and a designed frame.

Action: give movement a start and an end

Verbs describing a completed micro-movement render far more reliably than verbs describing a state. "She steps off the curb and turns toward the tram" supplies a start, a direction, and a destination. "She is waiting" supplies nothing to animate, so the model invents twitching to fill the gap.

One action per shot is the working rule. Two competing actions split the model's attention and usually produce both performed badly. If a shot needs three beats, it is probably three shots.

Environment: a floor, a horizon, something to catch the light

Environment is where physics gets anchored. Name the surface, the time of day, and one atmospheric detail: "rain-slicked Lisbon street at dusk, tram cables overhead, reflections in the cobblestones." Now the model knows where the ground is, where light comes from, and what should catch it.

Time of day is the cheapest lighting instruction available, so spend it deliberately. Blue hour, hard midday, overcast morning, sodium-lit night — each is shorthand for an entire lighting rig.

Camera and lens: turn a scene into a shot

Camera language is the highest-leverage part of a video prompt because it converts a description into a composition. Specify four things: framing (wide, medium, close-up), angle (eye level, low, overhead), lens character (24mm wide, 85mm portrait compression, macro), and one movement.

One movement per line. A push-in plus a pan plus a rack focus in the same sentence usually means none of them happen cleanly, because the model has to interpolate three simultaneous transformations across a handful of frames.

Light, palette, and texture

This is where mood lives. Name the quality and direction of light, then two color anchors, then a texture. "Soft north window light from camera left, desaturated teal shadows, warm skin tones, fine 35mm grain" is controllable. "Beautiful lighting" is a lottery ticket.

Texture words such as grain, haze, bloom, and halation tell the renderer how the image should degrade, and controlled degradation reads as expensive. Clean digital sharpness often reads as cheap because it exposes every synthetic seam.

Pace: how fast the feeling moves

Describe how fast things happen and how the frame should feel. "Slow, deliberate movement, measured walking pace, fabric drifting in light wind" is a timing instruction, not a mood word. Add a rig phrase when it matters: locked-off tripod, documentary handheld, smooth gimbal glide. Those phrases shape the micro-jitter that separates professional footage from accidental footage.

A reusable prompt skeleton

After twenty prompts, patterns emerge. This skeleton works because it moves from the most concrete information to the most stylistic, so early blocks are not competing with late ones.

[shot type + lens] of [subject with three visual details]
[one action with a start and an end]
in [environment + time of day + weather]
[light direction and quality], [two color anchors], [texture]
[one camera movement + pacing], [duration feel]
[optional audio or dialogue cue]

Two filled examples from very different briefs:

Medium close-up, 85mm compression, of a ceramicist in her fifties with
clay-dusted forearms and a linen apron, pressing a thumb into a bowl rim
in a sunlit studio at mid-morning, hard window light from camera right,
warm terracotta and dusty cream palette, fine 35mm grain, static camera
with a barely perceptible drift, slow quiet pacing
Wide establishing shot, 24mm, of a courier in a yellow rain jacket running
up wet stone steps in a coastal town at blue hour, streetlamps flickering on,
cool ambient light with warm practicals, deep blue and sodium orange palette,
light atmospheric haze, handheld follow from behind at running pace

Notice that neither prompt contains the word "cinematic." That word is a summary, not an instruction. Cinematic is what you get when framing, light, texture, and pacing are all specified.

One more habit worth building: put the camera clause on its own line. Instructions buried inside a long sentence get averaged away with everything around them. Instructions standing alone get obeyed.

Build the look before you ask for motion

Text is a lossy way to describe an image you can already see in your head. When a look matters, build the frame first: generate or source a still, iterate on composition and color while it is fast and cheap in the AI image generator, then use that still as the opening frame for image-to-video. This single habit removes most of the frustration from AI video work, because you stop asking a model to invent a look from adjectives.

Reference images do specific jobs, and knowing which job each reference performs keeps your prompt clean:

  • A face reference keeps identity stable across shots.
  • A wardrobe reference keeps a jacket the same jacket.
  • A lighting reference keeps a mood consistent through a sequence.
  • A compositional reference keeps the horizon at the same height.

Feed references alongside the text, not instead of it. The text still tells the model what to do with what it sees.

Audio behaves differently. If you can supply a music bed or an ambient loop, generate picture against it and let the rhythm inform your pacing cues. Even without an audio input, write the sound design into the prompt: ambient street noise, a single piano note, a voice at conversational volume. Motion designed for sound cuts better later, because the shot already has a rhythm to lock into.

If the tool supports lip-synced speech, keep spoken lines short, wrap them in quotation marks, and state who is speaking and to whom. Long monologues break sync and drift mid-sentence, and no amount of re-rolling fixes a structural problem.

Keeping a multi-shot sequence coherent

A sequence falls apart when shot three looks like a different production from shot one. Consistency is not a toggle; it is a small kit of artifacts you maintain across the whole project.

One canonical paragraph per character

Write a single paragraph for each character: age range, build, hair, one distinguishing feature, default wardrobe, posture. Paste it verbatim into every prompt where that character appears. Do not paraphrase between shots. Small wording changes produce small face changes, and small face changes are the most visible continuity error in generated video.

A style anchor line you never paraphrase

Create one line describing the film's look and reuse it unchanged: palette, light quality, lens family, grain. This line is your visual contract. If shot two drops the grain and warms the white balance, the audience reads it as a mistake rather than a choice.

Screen direction, prop state, and eyelines

Note what each shot hands to the next: a prop still in frame, the same direction of travel, a matching eyeline. Continuity in generated video is mostly about screen direction and object state. A cup that is full in shot two and empty in shot four without a reason is a break the model will happily produce for you.

Version logs and saved settings

When a render works, save the full prompt, the seed if your tool exposes one, and the reference set. Re-rolling rarely reproduces a lucky result, so a locked combination is worth more than a fortunate accident. Keep version notes like "S03 v4 — reduced haze, static camera, correct jacket" so you can rebuild the look weeks later when a client asks for a variant.

Camera and pacing vocabulary in plain language

You do not need a film degree, but you do need a small vocabulary. These phrases do most of the work:

  • Static or locked-off — no camera movement; the subject moves inside the frame. The safest choice for dialogue.
  • Slow push in — the camera advances, building pressure. Good for reveals and emotional beats.
  • Dolly out — the camera retreats, isolating the subject inside the environment.
  • Tracking or follow — the camera travels with the subject. State whether it follows from the front, side, or behind.
  • Orbit — the camera arcs around the subject. Keep it under a quarter turn or faces warp.
  • Crane up or tilt down — a vertical reveal, useful for scale and geography.
  • Handheld — deliberate instability. Specify intensity, or you get earthquake.
  • Rack focus — attention shifts between foreground and background. Powerful, easy to overuse.

Pacing deserves its own clause because it governs how a movement reads. A push-in across four seconds feels like tension; the same push-in across one second feels like a zoom. Say which one you want, and say it in the same breath as the movement.

Guardrails, negatives, and the contradictions they create

Negative prompts are the emergency brake, not the steering wheel. They suppress recurring artifacts but cannot repair a vague positive prompt. Keep the list short and specific to problems you have actually seen in your own renders.

Common entries: text overlays, watermark, logo, extra fingers, distorted hands, warped faces, duplicated limbs, sudden morphing, exposure flicker, oversaturated colors. Add "no cuts" when you want one continuous take, and "no slow motion" if the model keeps defaulting to it.

One caution: negatives can fight positives. Ask for a crowded market and then negate "extra people," and you have handed the model a contradiction it will resolve in the least flattering way. Review the whole prompt as a single instruction set rather than two independent lists.

The same logic applies to over-specified prompts. Every clause should remove a decision the model would otherwise make for you. If a clause removes no decision — "very nice," "high quality," "stunning" — delete it and use the space for a lens or a light direction.

Debugging: read the symptom before rewriting

When a render fails, diagnose before you rewrite everything. Most failures map to a handful of causes, and the fix is usually smaller than the frustration suggests.

Symptom Likely cause Fix
Faces melt mid-clip Too much simultaneous change Reduce action complexity, shorten duration, or split the shot
The camera ignores your move Instruction buried in a long sentence Put camera language on its own short line at the end
Everything looks flat and grey No light direction and no palette Add a named source and two color anchors
Unwanted text appears Model pattern-matching signage Remove sign, poster, label, or shop from the positive prompt; add text to negatives
The clip looks like a slideshow No cadence described Add pacing language and one clear movement
Style shifts between shots The style anchor line was paraphrased Paste it verbatim
Subject drifts off-model Missing or contradictory character description Restore the canonical paragraph and add a face reference
Motion stops halfway Duration exceeds what the description supports Shorten the clip or add a second beat as a separate shot

Keep your own failure list. A short personal record of what a specific model does wrong is worth more than any generic guide, because it is calibrated to your tool and your subject matter.

From beat sheet to first cut: the working order

  1. Write the beat, not the prompt. One sentence per shot describing what the audience learns or feels. If a shot teaches nothing, cut it before you spend a render on it.
  2. Build look frames first. Generate stills for the two or three key shots and lock the palette while it is cheap to iterate.
  3. Draft the continuity kit. Character sheet, style anchor line, prop list, screen direction notes. Ten minutes here saves an hour later.
  4. Fill every block of the skeleton on the first pass, even the ones you suspect do not matter. You learn which blocks your project is sensitive to only by specifying them.
  5. Generate the hardest shot first. If the difficult shot works, the rest is downhill. If it fails, you have not burned your energy on easy shots.
  6. Review at full speed, not frame by frame. Play clips at normal speed and judge whether they read. Micro-artifacts matter less than whether the motion tells the story.
  7. Assemble, then repair. Cut the sequence, find the shots that break the rhythm, and re-prompt only those. Start from the AI video generator for the reshoot rather than rebuilding the whole set.

Re-roll, repair, or rebuild

Experienced users spend less time generating than beginners because they know which failure they are looking at. Three questions sort most situations.

Is the failure compositional or incidental? If framing, subject, and light are right and only an artifact is wrong, re-roll. If the composition itself is wrong, repair the prompt.

Is the failure local or systemic? One bad clip in a good sequence is local — regenerate it. Every clip drifting toward the same problem is systemic — fix the style anchor or the character sheet before generating anything else.

Is the shot too ambitious for one generation? If it combines two characters, dialogue, a camera move, and a costume change, the honest answer is usually yes. Split it into an establishing wide, a coverage medium, and a reaction close-up. Three simple shots almost always beat one overloaded shot, and they cut better.

A useful budget rule: allow three attempts per prompt before you change the wording. Three failures of the same prompt is information; ten failures is stubbornness.

Mistakes that make generated footage look cheap

Describing a mood instead of a setup. Words like epic, dramatic, and beautiful are conclusions the audience draws. Build the conditions that produce them.

Writing a novel. Long prompts dilute. Details placed late compete with details placed early, and the model quietly drops constraints. Forty to ninety words is the practical range for most generators; go longer only when every clause removes a decision.

Changing five things at once. When a render fails, change one variable per attempt, or you will never learn which change fixed it.

Ignoring duration. A five-second clip forgives a loose description because there is little time to drift. A ten-second clip with two characters and a camera move needs tighter instructions. Match prompt precision to runtime.

Forgetting the edit. Clips that look fine alone can clash in sequence. Review them side by side, at identical size, in final order.

No naming convention. Untitled renders pile up and the good one disappears. Name files by shot and version from the first generation.

Treating a successful prompt as disposable. Prompts are assets. Keep a simple table with columns for shot ID, prompt version, references, seed, and verdict. That table will outperform memory every time.

Brand work deserves the same rigor. Capture a client's palette, visual habits, and lighting preferences in a reusable style block that lives beside the character sheet. That block is the difference between content that looks like it came from one studio and content assembled from five freelancers. Schedule a review pass across all renders in sequence before delivery, and fix the two worst offenders rather than the ten smallest.

Frequently asked questions

How long should a video prompt be? Long enough to specify subject, action, camera, light, and pacing, and no longer. For most generators that lands between forty and ninety words. Past that, later details start competing with earlier ones and the model resolves the conflict silently.

Should I include the word cinematic? It does little on its own. Replace it with the ingredients that create the impression: a specific lens, directional light, a defined palette, grain or haze, and controlled pacing.

Why does my character change between shots? Almost always because the description changed. Paste a canonical character paragraph verbatim into every prompt and add a face reference if your tool accepts one. Pair it with one unchanged style anchor line.

Do negative prompts fix bad output? They suppress recurring artifacts but cannot compensate for a vague positive prompt. Fix the description first, then add two or three targeted negatives for problems you have actually observed in your own renders.

How do I get smooth camera movement? State one movement, name its speed, and place it in its own short sentence. Keep the subject's action simple so the model is not solving motion and camera simultaneously. If it still jitters, shorten the duration by a second or two.

Can I reuse a prompt that worked on another project? Yes, and you should. Save successful prompts as reusable blocks — a lighting block, a camera block, a texture block. The skill is recombination, not writing something new every time. The prompt library is a useful place to study structural patterns before adapting them to your own subject.

What about two characters talking? Keep lines short, name the speaker, and use a static or slow push-in camera. Two people plus dialogue plus a moving camera is the hardest combination in AI video, so simplify one of the three variables before you generate.

How many shots should I generate before editing? Roughly one and a half times the number you plan to use, so the cut has options. Generating exactly the shot count forces you to keep clips that should have been replaced.

Is image-to-video always better than text-to-video? Not always. For a quick concept or a texture plate, text alone is faster. For anything where the look carries meaning — a product, a face, a brand palette — building the still first is faster in practice because it removes the guesswork.

Ready to direct your first shot? Open the AI video generator with the skeleton beside you and rebuild the frame block by block until it looks intentional rather than approximated. If you want a starting structure before you write your own, browse the template collection and treat each one as a working example of a specified shot rather than a finished style.

Then keep a log. The record of what worked, what drifted, and which line fixed it is what will make your next sequence look directed — not any list of magic phrases, but the accumulated evidence of your own renders.