Orelon logoOrelon
Precios

Prompt Writing for AI Image Generators: Cinematic Guide

5 nov 2026 · Por Orelon Team

Explora plantillas de video con IA

Echa un vistazo a algunas creaciones de la comunidad para inspirarte y abre cualquier plantilla para seguir creando en Orelon.

Learn how to write prompts for AI image generators that hold up in motion: structure, lighting, camera language, iteration loops, and fixes.

A good prompt reads like a shot list, not a wish. Most people start with adjectives — beautiful, cinematic, ultra-detailed — and get back images that look expensive and say nothing. The fix is structural. Describe what is in frame, how it is framed, how it is lit, and what changes between the first and last second of the shot.

That last part is where still-image habits break. A prompt tuned for a poster often collapses the moment you ask for motion, because motion needs verbs, camera behavior, and a duration to fill. This guide lays out a repeatable prompt method you can use for AI image generation and then carry straight into video work in Orelon's AI video generator. The skeleton is tool-agnostic: the same language works for a product still, a character portrait, or a four-second establishing shot.

Why image prompts and video prompts are not the same problem

A still image is a single decisive moment. The model only needs to resolve one frame, so ambiguity is cheap — if the pose is vague, you still get a usable pose. A video prompt has to resolve a moment and a trajectory. If you write "a woman walking through a market," you have given the model a subject and a location but no direction, no pace, no framing, and no reason for the camera to be anywhere in particular. The result is usually a drifting, vaguely handheld shot that looks like stock footage from nowhere.

What a still prompt optimizes for

Stills optimize for composition and legibility at a glance. You want the subject readable in under a second, the light doing something specific, and the palette coherent. Adjectives help here because they compress a lot of visual information: "overcast," "high-key," "desaturated teal" all push the model into a recognizable register.

What a shot prompt must add

Video prompts need three extra layers: a stated action with a beginning and end, a camera instruction that describes where the frame goes, and a sensory anchor that keeps the look stable across frames — usually light direction plus a texture word. Without those, the model invents continuity on its own, and continuity invented per frame looks like flicker.

The six-part prompt skeleton

Every prompt you write can be built from six slots, in roughly this order. Short prompts can skip slots; long prompts should never skip subject or light.

1. Subject and action

Name the thing, then give it something to do. "A cyclist" is a subject. "A cyclist coasting downhill, hands off the bars, jacket snapping" is a subject with action, and action is what gives a video model something to animate. Keep the count of moving elements low — one primary action, one secondary detail at most.

A ceramicist pressing a thumb into wet clay, shoulders tense, clay spinning steadily.

2. Framing and lens

Say where the camera is and what it is looking through. Shot size, angle, and lens character do more for perceived production value than any style keyword. "Wide, low angle, 24mm" tells the model to exaggerate the foreground. "Medium close-up, 85mm, shallow depth of field" tells it to separate the subject from the background.

3. Light and atmosphere

Light is the single highest-leverage slot. Naming the source and direction — "hard afternoon sun from camera left," "single practical lamp behind the subject" — produces more consistent results than naming a mood. If you want haze, fog, dust, or rain, say so, because atmosphere is what makes a frame feel like a place rather than a render.

4. Palette and texture

The model needs a material reality to hang the image on. Words like "matte," "brushed steel," "sun-bleached linen," or "film grain" do double duty: they set color and they tell the renderer what surfaces should feel like. Two texture words are usually enough; five make the frame noisy.

5. Motion and camera behavior

This is the slot still-image prompts leave out. Choose one camera move — slow push-in, lateral truck, static locked-off, gentle orbit — and one subject motion. Two moves fighting each other reads as chaos. If the shot is meant to be still, say "locked-off camera, no movement" so the model does not add drift.

6. Constraints

End with what you do not want, phrased positively where possible. "Clean background, no text, no watermark, single subject" prevents the most common failures. Negative phrasing works better as a short list than as a paragraph.

A worked example, three levels deep

Here is the same idea written at three levels of detail. Notice that the strong version is not longer because it has more adjectives — it is longer because it has more decisions.

Weak: A cozy cabin in the woods, cinematic, beautiful lighting.

Medium: A small wooden cabin at dusk, warm light in the windows, pine forest around it, mist near the ground, wide shot, cinematic.

Strong: Wide establishing shot of a single-room timber cabin at blue hour, warm tungsten glow spilling from two windows onto wet gravel, low ground mist drifting left to right, dense pine treeline behind, 28mm lens, slight barrel distortion, camera locked off, cool ambient blue against warm interior, fine film grain.

The strong version can be converted to motion with one line: slow dolly forward toward the door over four seconds, mist continues drifting left to right. Everything else stays. That is the whole trick — write an image prompt detailed enough that adding a camera move is the only remaining decision.

Camera vocabulary that actually changes output

Vague camera words get ignored. Specific ones get rendered. A short reference table helps when you are writing quickly:

Intent Prompt phrase What it changes
Establish scale wide shot, 24mm, low angle Exaggerates foreground, makes space feel large
Isolate a face medium close-up, 85mm, shallow depth of field Compresses background, softens edges
Create intimacy handheld, slight breathing motion Adds imperfection and presence
Show detail macro, 100mm, hard raking light Reveals surface texture
Signal tension slow push-in, static subject Movement into stillness reads as threat

Longer lens descriptions are useful when you are chasing a specific look — anamorphic flares, vintage glass softness, or the flattened perspective of a long telephoto. Keep them to one per prompt; stacked lens words average out into mush.

Lighting is the cheapest quality upgrade

If you only fix one slot, fix light. Three properties matter: direction, quality, and color temperature. Direction tells you where shadows fall. Quality (hard versus soft) tells you how edges behave. Color temperature gives you the emotional register — cold blue for unease, warm amber for safety, green for institutional.

A prompt that says "golden hour" gives you a time of day. A prompt that says "low sun from behind camera right, long shadows raking across the floor, warm 3200K against cool shade" gives you a lighting plan. The second one produces frames that look like they were shot rather than sampled.

For video specifically, keep the light plan static unless the shot is about light changing. A moving light source forces the model to re-solve the frame every second, which is where a lot of shimmer comes from.

Debugging a bad generation: a five-step loop

When output is wrong, resist rewriting everything. Change one variable at a time so you learn something.

  1. Check subject clarity. Can a stranger describe the frame in one sentence? If not, simplify the subject and action before touching style.
  2. Check framing. If the composition feels generic, add shot size and lens. Generic framing is the most common cause of "AI-looking" output.
  3. Check light. If the image is flat, add direction and quality. If it is muddy, remove one atmosphere word.
  4. Check motion. If the video drifts or warps, cut secondary actions and lock the camera.
  5. Check constraints. If artifacts persist, add them to a short negative list at the end rather than scattering "no" clauses through the prompt.

Run three variations at a time and compare, rather than generating twenty near-identical frames. Three deliberate variants teach you more than twenty random ones.

Prompt patterns by use case

Product shots that stay on-model

Describe the object as if briefing a photographer: surface finish, edge treatment, and how light wraps it. "Matte ceramic bottle, soft key from above, subtle rim light on the right edge, seamless pale grey background, 100mm macro, no reflections of the studio." For video, add a single slow orbital move and keep the background plain so the model has nothing extra to track.

Character consistency

Consistency comes from a fixed block of text you reuse verbatim — face shape, hair, wardrobe, and one signature detail — followed by a variable block for pose and setting. Reordering or rewording the fixed block between generations is the number one reason characters drift.

Environments and establishing shots

Establishing shots benefit from a foreground, a middle ground, and a background element that implies depth. "Wet cobblestones in the foreground, a figure with an umbrella at mid-distance, cathedral silhouette in haze beyond." That three-layer structure gives a camera move something to travel through.

Abstract and title sequences

Abstract work is where texture words earn their keep. Pick one material behavior — ink dispersing, glass refracting, smoke folding — and one motion, then resist adding more. Abstract prompts fail by accumulation, not by lack of detail.

Common mistakes that quietly wreck output

  • Stacking style keywords. "Cinematic, hyperreal, 8k, award-winning" pulls in conflicting references. Pick one register.
  • Contradictory camera instructions. A locked-off dolly shot is not a thing. Choose.
  • Too many subjects. Every additional moving subject splits the model's attention and increases warping.
  • No duration sense. In video prompts, mention pace — "slow," "over four seconds," "unhurried" — so motion does not accelerate.
  • Ignoring aspect ratio. A vertical prompt composed like a landscape frame loses its subject. Compose for the format you will deliver.
  • Rewriting the whole prompt after one bad result. You lose the information about what actually worked.

Build a reusable prompt system instead of one-off prompts

Professionals do not write prompts from scratch each time. They keep a small library of tested blocks: three lighting setups, four lens and framing combos, two motion patterns, one negative list. New work is assembled from proven parts, which is why their output looks consistent across a campaign.

Start with a document split into those four categories, and append a new block only when a generation genuinely earns it. Keep a note of the exact wording — paraphrasing a block usually changes the output. If you want a running start, browse the Orelon prompt library for structure examples, then adapt the skeleton above to your own visual language. For teams producing volume, pairing a stable prompt library with reusable video templates removes most of the setup cost per asset.

The workflow that scales looks like this: generate the still in the AI image generator until composition and light are right, then animate that approved frame rather than prompting video from scratch. Image-first pipelines are more predictable because you approve the frame before motion introduces new variables.

FAQ

How long should a prompt be?

Long enough to cover subject, framing, light, texture, motion, and constraints — usually 40 to 90 words. Beyond that, extra words tend to dilute rather than refine. If you need more control, split the idea into multiple shots instead of one enormous prompt.

Should I use the same prompt for image and video?

Use the same foundation, then add motion. The image prompt defines the frame; the video prompt defines what happens inside it and where the camera goes. Adding motion language to a strong still prompt is far more reliable than writing a video prompt from nothing.

Why does my video look warped even though the image was fine?

Usually too much simultaneous movement. Cut secondary actions, lock the camera, and simplify the background. Models handle one clean action well and struggle with layered motion, especially at the edges of frame.

Do negative prompts help?

Yes, but keep them short and specific — text, watermark, extra limbs, duplicate subject. Long negative lists start removing things you wanted, because the model treats them as concepts present in the scene.

How do I keep a character consistent across shots?

Freeze a description block and reuse it word for word, including the odd detail. Change only pose, framing, and setting between shots. Consistency is a copywriting discipline more than a technical setting.

What is the fastest way to improve my results?

Add shot size and light direction to every prompt for a week. Those two slots account for most of the difference between an image that looks generated and one that looks directed.

Put the prompt to work in motion

Prompting is a craft of decisions, not adjectives. Write the subject with an action, place the camera, name the light and its direction, anchor the texture, then add exactly one motion. Debug one variable at a time, and keep the blocks that work.

When your frame is ready to move, take it into Orelon and turn a still into a cinematic shot. Start from an approved image, describe the camera move and the pace, and let the idea carry the motion — because a prompt is only as good as the shot it produces. Explore the AI video generator to test the method on your own footage-style concepts, and compare approaches in the Orelon blog when you want a second workflow to pressure-test against.