Orelon logoOrelon
Precios

AI Text-to-Image Prompting: Write Prompts That Work

7 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 structure AI text-to-image prompts, adapt them for motion, keep style consistent across shots, and build a workflow from mood board to final clip.

Most disappointing AI images are not caused by weak models. They are caused by prompts that describe a mood instead of a picture. "A lonely woman in a city at night, cinematic, beautiful lighting" gives the model a hundred valid interpretations and no reason to choose one. A prompt that names the lens, the light source, the wardrobe, the weather and the frame shape leaves far less to chance — and that difference shows up long before you start swapping tools.

This guide treats text-to-image prompting as a working craft skill: how to structure prompts so they survive iteration, how to carry them into motion, and how to build a workflow that keeps a look consistent from the first still to the final cut.

What a prompt generator actually does

A prompt generator is not a magic button that turns a vague idea into a masterpiece. It is a text tool that does three specific jobs, and knowing which one you are using changes how much you get out of it.

Job one: expansion

You type "rainy Tokyo alley, neon, moody" and the tool returns a paragraph mentioning wet asphalt reflections, steam from a vent, magenta and cyan practicals, a 35mm lens, shallow depth of field, and a 2.39:1 frame. Expansion is useful when you have a feeling but no vocabulary. It hands you the descriptive words you were missing.

Job two: structuring

Better generators impose an order: subject, action, environment, light, camera, style, constraints. Structure matters because diffusion models weight early tokens more heavily in many implementations, and because a consistent order makes your own prompts easier to debug. If a result is wrong, you can scan the structure and find which slot failed instead of rewriting the whole thing blindly.

Job three: translation

Different models respond to different dialects. One model likes comma-separated keywords and responds well to photographic jargon; another prefers full sentences and tends to ignore stacked modifiers. A generator that produces variants for different systems saves you from re-learning phrasing every time you change tools.

What it cannot do for you

It cannot know your intent. It does not know that the character must wear the same jacket in shot four as in shot one, that the client hates orange, or that the scene needs to read as winter rather than autumn. It also cannot judge whether an output is good — that stays entirely with you. Treat generated prompts as a first draft written by a fast assistant with no taste.

Anatomy of an image prompt that survives iteration

Every strong image prompt answers the same handful of questions. You do not need to answer all of them every time, but the more you answer, the more control you keep.

Subject and action

Name who or what, and what they are doing. "A woman" is a category. "A woman in her thirties mid-stride, coat half-open, looking over her shoulder" is a picture. Action gives the model a body to pose and a direction to move.

Environment and time of day

Place the subject: a bus shelter, a rooftop water tank, a hospital corridor. Then anchor the light with a time of day or a weather condition. "Late afternoon" and "overcast noon" produce completely different images even with identical subjects.

Light — the biggest single lever

If you change only one thing in a prompt, change the lighting description. Useful vocabulary:

  • Source: window light, practical lamps, neon signage, firelight, bounced flash, hard sun.
  • Quality: soft, hard, diffused, specular, dappled.
  • Direction: side-lit, backlit rim, top-down, underlit.
  • Ratio: high key, low key, Rembrandt, split.

"Backlit by a single tungsten bulb, dust in the air, deep shadow on the face" will outperform any number of vague adjectives like "dramatic" or "moody."

Camera and lens

Camera language is shorthand for a whole set of visual decisions. A 24mm lens implies a wide, slightly distorted space with deep context. An 85mm at f/1.4 implies compression and a melted background. A macro lens implies texture. Naming the lens plus the aperture plus the distance gets you three decisions for the price of one.

Medium and grade

Decide whether you are making a photograph, a still from an animated film, a watercolor, a risograph print or a 3D render. Then describe the color treatment: warm highlights and cool shadows, bleach bypass, muted pastels, high-contrast black and white. Medium and grade together define the texture that makes a set of images feel like one project.

Constraints

Constraints are the least glamorous slot and the most underused. They include the aspect ratio, what must not appear in frame, and how much empty space you need for titles. If a clip will carry a caption, ask for negative space at the bottom and you save yourself an edit later.

The six-slot formula in practice

Here is a compact, reusable structure:

Subject + Action | Environment + Time | Light | Camera + Lens | Medium + Grade | Constraints

Filled in:

A courier in a soaked raincoat, checking a paper address | an empty riverside loading dock at dusk | single overhead sodium lamp, wet concrete bounce | 40mm, eye level, shallow focus | 35mm film look, cool shadows, faint grain | 16:9, no visible text, space on the left

That prompt is not poetry, and it does not need to be. It is a set of instructions with enough specificity that two different models will produce recognizably the same scene.

Note what is absent: no "8k", no "masterpiece", no "award-winning". Those tokens are mostly noise in modern systems. Specific nouns and real photographic terms do the work.

From still to motion: adapting prompts for video

Video prompts inherit everything from image prompts and then add two new dimensions: change over time, and camera behavior. This is where most people's results collapse.

Describe change, not just appearance

A video model needs to know what moves. Say it plainly: hair shifting in wind, steam curling from a cup, a curtain settling, the subject turning from profile to three-quarter. If nothing is described as moving, the model will often invent small, meaningless motion to fill the frame — the subtle warping that makes clips feel artificial.

Camera vocabulary that models understand

  • Dolly in / push in: builds focus and intensity.
  • Tracking shot: camera follows the subject laterally.
  • Crane up: reveals scale.
  • Handheld follow: adds tension and immediacy.
  • Static locked-off frame: lets performance carry the shot.
  • Rack focus: moves attention between foreground and background.

Pick one primary move per shot. Two competing camera instructions in a single prompt usually produce drift.

Timing and shot length

State the duration you are aiming for and what should happen inside it. "Three seconds, she turns her head once, no other motion" is a real instruction. Vague prompts on long durations invite the model to fill time with morphing.

Shorter clips with clearly defined single actions cut together far better than long clips full of improvisation.

A workflow that keeps quality high and wasted effort low

Step 1: Collect references before you type a word

Gather six to ten images that capture the light, palette and texture you want. Write down what you actually like about each one in plain language: "the way the window light wraps the shoulder," "the green in the shadows." This list becomes your prompt vocabulary. Prompting without references is guessing.

Step 2: Write a master style block

Your style block is a paragraph you paste into every prompt in a project. It holds the medium, grade, grain, lens family and any recurring environmental detail. Keeping it identical across shots is what makes separate generations feel like one film.

Step 3: Generate stills first

Stills are cheaper, faster and easier to judge than video. Use an image workflow to nail the look — light, palette, wardrobe, framing — before committing to motion. Approving a look on a still and then animating it costs a fraction of animating five wrong versions. The AI image generator is the natural place to run this exploration loop.

Step 4: Lock the look, then animate

Once a still works, use it as the first frame and write a motion prompt that describes only what changes. This split — appearance in the image prompt, change in the motion prompt — is the single most reliable way to get controlled video output. You can move from approved frames to clips in the AI video generator without rewriting the visual description.

Step 5: Edit in sequences, not single shots

A list of beautiful clips is not a scene. Build sequences: wide to establish, medium to orient, close to emote, insert for texture. When you prompt, ask yourself which role each shot plays, and set the framing accordingly. Consistency between shots matters more than the brilliance of any one frame.

Consistency controls worth mastering

  • Locked style block. One paragraph, reused verbatim, across the whole project.
  • Character sheet. Front, three-quarter and profile descriptions with fixed wardrobe, hair and age wording. Paste the same wording everywhere.
  • Seed reuse. Where the tool exposes a seed, keep it fixed while you adjust one variable at a time.
  • Reference images. Feeding a reference frame is usually more powerful than a hundred extra adjectives.
  • Naming discipline. Name files with scene, shot and version. Consistency dies in unnamed folders.
  • Deliberate variation. Change one slot per iteration. Two changes at once and you learn nothing about which one worked.

Common prompt mistakes and quick fixes

Mistake What it looks like Fix
Adjective stacking "stunning, epic, beautiful, cinematic" Replace each adjective with a concrete noun or light source
No camera decision Random framing per generation Name lens, height and distance
Contradictory light Sunlight and neon and moonlight together Choose one dominant source plus one fill
Motion left unstated Warping and drift in video Describe the single action and the camera move
Style drift Each shot looks like a different film Reuse one style block and one palette description
Overlong prompts Late instructions ignored Move constraints to the front or trim the middle

Quality control: a sixty-second checklist

Before you accept an output, check:

  1. Anatomy. Hands, eyes, teeth, ears. Small errors compound when animated.
  2. Text. Any signage or labels should be removed or reframed, not tolerated.
  3. Physics. Fabric folds, hair weight, liquid behavior. Wrong physics reads as cheap.
  4. Temporal stability. Watch for flicker in backgrounds and edges of the frame.
  5. Continuity. Compare against the previous shot for palette, light direction and wardrobe.
  6. Usability. Is there space where the edit needs it? Is the framing crop-safe for vertical delivery?

If a shot fails three or more checks, do not patch it in the edit. Regenerate.

Where prompt libraries and templates fit

Once you have written a style block that works, save it. The compounding value of prompting is not in any single prompt but in the reusable components: a lighting phrase that reliably produces soft window light, a lens phrase that reliably produces compression, a motion phrase that reliably produces a slow push without drift.

A curated prompt library shortens the distance between an idea and a usable draft, especially when you are moving between genres — product beauty shots, documentary texture, stylized animation. The same logic applies to video templates: they encode structural decisions like pacing and shot rhythm so you can spend your attention on the parts that are genuinely yours.

How to know a prompt is working

A good prompt has a signature: it produces variations on the same idea rather than a lottery of unrelated images. Test yours by running it three times. If the three results share a palette, a light direction and a framing family, the prompt is doing its job. If they could belong to three different projects, the prompt is under-specified — usually in light and camera.

You can read more about structuring creative workflows on the Orelon blog.

FAQ

Do prompt generators replace prompt writing skill?

No. They accelerate the first draft and supply vocabulary, but the decisions that matter — intent, references, constraints, judgment about what is good — remain yours. The strongest results usually come from people who use a generator for expansion and then edit the output ruthlessly.

How long should a text-to-image prompt be?

Long enough to cover subject, environment, light, camera, style and constraints, and no longer. In practice that is often 30 to 60 words. Beyond that, later instructions start getting ignored, especially in video models with limited prompt attention.

Should I use negative prompts?

Yes, but sparingly. A short list of things you genuinely do not want — visible text, extra limbs, heavy watermark artifacts, lens flare — helps. A long list of stylistic dislikes usually backfires, because naming a thing often summons it.

Why do my video clips morph and warp?

Usually because the prompt describes a static scene and leaves motion undefined. State one action and one camera move, keep the clip short, and start from a strong still frame. If the still is ambiguous, the animation will amplify the ambiguity.

Can I use the same prompt across different models?

Broadly yes, but expect a translation step. Photographic jargon travels well; stacked quality keywords and platform-specific syntax do not. Keep a core description you trust and adjust the modifiers per system.

How do I keep a character consistent across many shots?

Write a fixed character description once, paste it verbatim, pair it with a reference image, and keep the style block identical. Where seeds are available, keep them stable while you change only the framing.

Is it worth learning lighting vocabulary?

It is the highest-return vocabulary you can learn. Terms like backlit, rim light, softbox, practicals and bounce give you direct control over the part of an image that viewers read first.

Start with one prompt you can repeat

The fastest way to improve your output is not a new model. It is a prompt structure you trust, a style block you reuse, and the discipline to change one variable at a time. Write the six slots, approve the still, then describe the motion.

Put that loop to work: explore the look in the AI image generator, then bring the approved frame into the AI video generator and give it one clear action and one clear camera move. If you want to see how a finished sequence should read, browse the Seedance 2.5 examples for motion reference, and check Orelon pricing when you are ready to scale a project. Cinematic ideas in motion start with a sentence precise enough to be filmed.