Learn how to write image prompts that make AI video feel cinematic, with framing, light, motion cues, continuity checks, and a repeatable shot workflow.
A video model rarely fails because it lacks capability. It fails because the first frame was vague. Ask an AI video generator for a cinematic shot of a city and you may get something that moves but says nothing. Give it a frame with a clear subject, a deliberate lens, a specific light source and an implied direction of travel, and the same tool can return a clip you would actually cut into a timeline.
That is the craft of image prompting for video: turning a feeling into a visual specification precise enough to reproduce, yet open enough to let the model do interesting work. This guide covers how to build that specification, how to test it, how to keep a look consistent across a sequence, and how to avoid the small errors that make generated footage feel like a demo instead of a scene.
The first frame is the brief
When you feed an image into a video model, the still becomes ground truth. Composition, palette, lens character and subject identity are largely locked the moment the frame is accepted. Text prompts then steer behaviour: how the camera moves, how fast the subject walks, whether the light flickers or holds.
That split is the most useful mental model in AI filmmaking. The image is the what. The prompt is the how. Creators who fight this usually write long text descriptions and hope a random still emerges. Creators who work with it build the frame deliberately, then describe motion in a separate, shorter pass.
A practical consequence: never try to fix a badly composed still with motion instructions. If the horizon is tilted, the subject is centre-frame and there is dead space on both sides, no amount of slow dolly-in will rescue it. Regenerate the frame instead. A new still costs seconds; a weak clip costs a review cycle, a re-edit and sometimes a whole re-shoot in AI terms.
This is also why image prompts deserve more attention than they usually get. Most tutorials spend their energy on camera moves and overlook the fact that the move only has something to reveal if the frame is worth exploring. A strong still with a simple push-in will outperform a weak still with a complex crane move every time.
Anatomy of an image prompt that moves
A strong video-ready image prompt has five layers. You do not need all five in every prompt, but knowing which layer is missing tells you why a result feels flat. The layers are subject and action, lens and framing, light and time of day, palette and texture, and format with motion space.
Subject, action and intent
Name the subject specifically and give it one action in progress. A woman in her thirties turning to look over her shoulder is workable. A woman is not. The verb matters more than the adjective: models render posture, weight and direction of gaze far more reliably than they render personality.
Add intent when the story needs it. A courier checking a phone at a locked door implies a narrative. The same courier standing in a corridor implies nothing. You do not need a full backstory in the prompt, but you do need a situation that a viewer can read in under a second.
Lens, framing and camera height
Lens language is the fastest way to make AI output read as cinema rather than stock. Use focal lengths the way a cinematographer would: 24mm for environmental context, 35mm for handheld intimacy, 50mm for neutral observation, 85mm for compressed portraits. Pair each with a camera height: eye level, chest height, low angle looking up, overhead.
The combination does more than any adjective. A 24mm frame from chest height feels present and slightly vulnerable. An 85mm frame at eye level feels intimate and compressed. A 35mm frame from a low angle feels confrontational. Decide what the shot needs emotionally before you decide what it needs visually.
Composition rules still apply, but treat them as tools rather than laws. Placing your subject off-centre gives the frame breathing room and gives the model a natural direction for a pan or push. Leave negative space where the action will travel. If the subject is about to walk left, do not fill the left side with clutter.
Light source, quality and time of day
Lighting is where most prompts collapse into vagueness. Cinematic lighting means nothing on its own. Instead, describe the source and its quality: hard afternoon sun through blinds, soft north-facing window light, sodium streetlamps with wet reflections, a single practical lamp with the rest of the room in shadow.
Time of day helps, but it is not a substitute for a source. Blue hour tells the model the colour temperature; it does not tell the model where the light is coming from. Combine the two. A tired nurse steps out of a hospital service door into blue pre-dawn light, with a warm corridor spilling through the gap. That gives the model a direction, a contrast and a story beat.
Palette, texture and format
Describe the palette in three words or fewer. Teal shadows with warm skin. Desaturated grey-green with one red accent. Bleached highlights and deep browns. Restraint here is what makes a sequence feel graded rather than generated.
Texture phrases control the finish: fine 35mm grain, clean digital sharpness, slight halation on highlights, matte filmic contrast, mild vignette. Format phrases control the frame: 16:9 anamorphic, 9:16 vertical, 2.39:1 widescreen. Decide the delivery format at prompt time, not in the edit. A vertical composition cropped from a widescreen frame often loses its subject and its negative space.
Negative space and motion path
Finally, think about where motion will happen. If the camera is going to push in, leave room at the centre. If the subject is going to walk through frame, leave an entry point and an exit point. If steam or fabric will move, place it where the light can catch it.
This layer is often invisible in the still but obvious in the clip. A frame that is perfectly balanced as a photograph can become crowded the moment the camera moves. Build the still with the upcoming movement in mind.
A four-pass workflow from logline to clip
The strongest workflows separate still generation from motion generation, and they do it in four passes. The sequence is simple enough to repeat and structured enough to diagnose when something goes wrong.
Pass 1: write the logline as a visual sentence
One sentence, one subject, one place, one light. A tired nurse steps out of a hospital service door into blue pre-dawn light. That sentence becomes your prompt skeleton and later your shot list entry. If you cannot write the sentence, you are not ready to write the prompt.
Keep the sentence physical. Avoid abstract goals like show the feeling of exhaustion. Instead, choose an action that implies exhaustion: leaning against a doorframe, fumbling for a key, pausing before stepping into the cold. The model can render behaviour; it cannot render a mood label.
Pass 2: generate a small reference batch
Generate a small batch at the target aspect ratio using an AI image generator. Judge quickly and harshly on three criteria only: does the subject read instantly, is the light doing something, and is there space in the frame for the action you plan. Delete everything else. Ten seconds of judgment beats ten minutes of rationalisation.
Do not over-generate. If you make forty stills, you will pick the most polished one rather than the one that serves the shot. A batch of four to eight is usually enough to see whether the prompt is working. If none of them work, the prompt is missing a layer, not a lucky seed.
Pass 3: translate the still into motion language
Now write the motion prompt without repeating the image description. Two sentences at most: what the camera does, and what changes in the frame. Slow push in toward the doorway as her breath fogs. That is enough. Repeating the visual description in the motion prompt often causes the model to re-render details and drift.
Use motion prompts to answer a simple question: what is the one thing the viewer should notice changing? If everything changes, nothing lands. If nothing changes, the shot may still work as a tableau, but it will not feel alive.
Pass 4: extend in short blocks
Generate five seconds, review, then extend. Long single generations accumulate drift: faces soften, wardrobe changes, light shifts direction. Short blocks with manual review keep continuity under control and let you throw away only the weak section.
A prompt library helps here. Save the phrases that consistently work, such as lens descriptions, grain settings and palette triples, and reuse them instead of rewriting from memory. Orelon's prompt library is a reasonable starting point if you want proven phrasing rather than guesswork.
| Pass | Output | Review question |
|---|---|---|
| Logline | One visual sentence | Can I see the shot in my head? |
| Reference still | One or two candidate frames | Does the subject read and does the light work? |
| Motion prompt | One camera move and one change | Is the instruction short and physical? |
| Extension | Short reviewed blocks | Has identity or light drifted? |
Motion cues that work with the frame
Motion is not a substitute for composition. It is a way to reveal composition over time. The best motion prompt usually describes a simple physical action that the frame was designed to support.
Camera motion vocabulary
Use a small vocabulary and use it consistently. Slow push in. Lateral tracking left. Gentle handheld drift with subtle breathing. Static frame with the subject walking through. Slow tilt up from hands to face. Rack focus from foreground label to product. These phrases are understandable, physical and easy to review.
Avoid stacking moves. A push-in with a simultaneous orbit, zoom and tilt will usually produce mush because the model has to reconcile conflicting instructions. Choose one primary move. If you need a second idea, make it internal to the frame rather than another camera instruction.
Internal motion is often stronger
If you want a shot to feel alive without moving the camera, describe what moves inside the frame: steam rising from a cup, fabric flexing, rain streaks catching light, hair shifting in wind, a curtain breathing at an open window. Internal motion is often more convincing than camera motion because it respects the stillness of the composition.
This is especially useful for portraits and product shots. A static frame with a subtle shift in expression or a highlight moving across a surface can feel more cinematic than a dramatic camera move that calls attention to the generation.
Motion conflicts and how to spot them
A motion prompt is probably too complex if it contains more than two verbs that affect the whole frame. Watch for instructions like the camera circles the subject while pushing in and the background blurs. Those are three post-production ideas compressed into one generation. Split them into separate shots or drop two of them.
Also watch for motion that contradicts the still. If the still is a tight 85mm portrait, a wide environmental pull-back will force the model to invent a world that was never in the frame. Design the move around the information the still already contains.
Continuity across a sequence
Single clips impress. Sequences sell. Continuity in AI video comes from repetition, not from memory. The model does not remember your earlier shot, so you have to re-specify the world every time.
The reusable style block
Create a short style block and paste it into every prompt in the sequence: lens, palette, grain, light direction and aspect ratio. Six to ten words total. Then vary only the subject and action lines. This is the same discipline a colourist applies across a scene, translated into text.
An example style block: 35mm, eye level, teal shadows, warm skin, fine grain, 16:9. That block does not describe the subject. It describes the visual rules of the world. Reuse it until the sequence is finished.
Character description order
For characters, always describe them in the same order: age, build, hair, wardrobe anchor, distinguishing detail. Changing the order changes what the model emphasises, and the face drifts. One wardrobe anchor, such as a grey wool coat or a scuffed canvas jacket, does more for recognisability than three paragraphs of biography.
Keep the anchor specific but not fussy. A red scarf is easy to reproduce. A hand-stitched indigo scarf with a small silver pin is harder, and the model may simplify it anyway. Pick one detail that survives compression.
The shot ledger
Track your shots in a simple table: shot number, duration, subject, action, camera move, notes. It sounds bureaucratic until you are twelve clips deep and cannot remember which one had the rain. The ledger also makes it easy to hand a sequence to an editor or collaborator.
| Shot | Duration | Subject and action | Camera | Notes |
|---|---|---|---|---|
| 1 | 5s | Nurse pushes service door open | Slow push in | Blue pre-dawn, warm corridor gap |
| 2 | 5s | Figure crosses wet lot | Static, slight handheld | Wider 24mm, door closing behind |
| 3 | 5s | Hands in pockets, steam crosses frame | Static | 85mm close-up, same grain |
Image-to-video or text-to-video
Both paths belong in the same toolkit. The decision is about control versus speed. Image-to-video is stronger when the frame itself is the contract: a client-approved look, a recurring character, a product whose geometry must stay intact. Text-to-video is stronger when you are exploring, testing a mood or building an abstract transition that does not need a literal anchor.
| Situation | Better path | Why |
|---|---|---|
| Client-approved look and brand colours | Image-to-video | The still is the visual agreement |
| Recurring character across shots | Image-to-video | Identity lock is stronger |
| Exploring an idea in five minutes | Text-to-video | Faster, with no still step |
| Abstract or surreal transitions | Text-to-video | Less literal anchoring |
| Product accuracy required | Image-to-video | Geometry stays intact |
| B-roll for a voiceover edit | Either | Volume matters more than polish |
A useful hybrid: use text-to-video for rough exploration, screenshot the frame you like, then re-enter it as an image prompt and regenerate with proper resolution and composition. You keep the discovery and lose the sloppiness.
Another hybrid is to build the reference still in an image tool, animate a short test, then return to the still and adjust one layer before generating the final clip. This keeps the visual decision separate from the motion decision. It also prevents the common mistake of endlessly regenerating video when the real problem is a weak frame.
Prompt patterns by deliverable
Different deliverables need different prompt shapes. These patterns cover most commercial work, and each can be adapted by changing the style block rather than rewriting the whole prompt.
Cinematic narrative shot
Subject in mid-action, 35mm lens, eye level, motivated practical light, shallow depth of field, fine grain, 2.39:1. Motion: slow push in, subject continues action. Emphasis on performance and light rather than spectacle. Use this when the shot needs to carry emotion or story information.
Product and packshot
Subject isolated on a defined surface, 85mm macro, controlled studio light with one rim highlight, clean shadows, matte finish, 16:9. Motion: slow orbital drift or a single rack focus from label to full product. Avoid busy backgrounds; reflections read as noise. If the product has text, keep the angle simple and avoid motion that smears the label.
Documentary and b-roll
Wide environmental frame, 24mm, available light, mild handheld imperfection, natural colour, 16:9. Motion: gentle drift or static with internal movement. Slight imperfection is a feature here. Too clean reads as advertising. Use this pattern for establishing shots, transitions and coverage under a voiceover.
Vertical social loop
Subject centred in the upper two-thirds, 35mm, bright even light, high contrast, 9:16. Motion: one continuous move that ends where it began so the loop is invisible. Design the first and last frame to match before you generate. If the loop point is obvious, the clip will feel like a stutter rather than a seamless moment.
Explainer insert or abstract transition
Macro detail, controlled light, shallow depth of field, clean background, 16:9. Motion: rack focus, slow lateral slide or a single element entering frame. Keep the idea small. An insert shot is not a montage; it is one visual fact that supports the narration.
Worked example: a three-shot sequence
Take the logline from earlier: a tired nurse steps out of a hospital service door into blue pre-dawn light. The goal is a short sequence that feels like the opening of a scene, not three unrelated clips.
Style block: 35mm to 85mm, eye level, blue pre-dawn with warm corridor spill, teal shadows, fine grain, 16:9.
Shot one still prompt: a nurse in her forties in scrubs pushing open a metal service door, seen from chest height, 35mm lens, blue pre-dawn light with a warm corridor spilling through the gap, wet asphalt reflections, off-centre composition, fine grain, 16:9.
Shot one motion prompt: slow push in as the door swings wider and her breath fogs in the cold air.
Shot two still prompt: wider 24mm frame from across the lot, the same nurse walking toward camera, the service door closing behind her, wet asphalt, blue pre-dawn light, a single warm window in the background, fine grain, 16:9.
Shot two motion prompt: static frame with a slight handheld drift as she walks toward camera.
Shot three still prompt: 85mm close-up of hands in pockets, steam from a vent crossing frame, blue pre-dawn light, faint warm rim on the knuckles, fine grain, 16:9.
Shot three motion prompt: static frame, steam drifting left to right, fabric shifting slightly.
Three clips, one look, about fifteen to twenty seconds of screen time. Assembled with a sound design pass, it reads as a scene rather than a demo because every prompt re-specified the same world instead of relying on luck. The shot ledger would note the lens, the light direction and the wardrobe anchor so a fourth shot could be added later without breaking the sequence.
Review, mistakes, and FAQ
A review loop is part of the craft, not an admission that the tool is unreliable. The faster you can reject a weak frame or clip, the more consistent your output becomes.
Mistakes that flatten a shot
- Stacking adjectives instead of specifying light. Replace beautiful, moody and epic with one light source and one quality word.
- Describing a montage in one prompt. One shot, one action. Sequences belong in the edit, not in a single generation.
- Ignoring the aspect ratio until export. Compose for the delivery format from the first still.
- Letting the camera do everything. A static frame with strong internal motion often outperforms a dramatic move.
- Skipping review between blocks. Drift is cheap to fix early and expensive to fix late.
- Changing character description order. Keep age, build, hair, wardrobe anchor and detail in the same sequence every time.
- Forgetting negative space. A frame that is full at rest has nowhere to move once the camera or subject starts moving.
Review criteria that save time
Check the still at thumbnail size first. If the subject does not read at thumbnail size, it will not read in motion. Then check the light: is there a clear direction and a clear contrast? Then check the action space: is there room for the planned movement? Finally, check the format: does the composition respect the target aspect ratio?
For clips, watch once without pausing. If you notice a technical error before you notice the subject, the clip needs work. Then watch again with the shot ledger open and compare identity, wardrobe, light direction and grain against the previous shot. Keep notes short and specific so the next prompt can fix the issue.
FAQ
How long should an image prompt be?
Long enough to specify subject, lens, light and format, usually 25 to 45 words. Beyond that, prompts tend to contradict themselves. If you need more control, add a reusable style block rather than lengthening individual sentences.
Why does my character change between shots?
Because nothing in the model persists between generations. Fix it by describing the character in an identical order every time, keeping one distinctive wardrobe anchor, and generating the reference still from the same prompt base. Expect to regenerate occasionally and budget attention for it.
Should I write motion prompts in the same message as the image?
No. Let the still establish the frame, then describe motion separately and briefly. Mixing a full visual description into the motion prompt encourages the model to reinterpret details it had already rendered correctly.
Do longer generations save time?
Usually not. Longer clips drift more and give you fewer decision points. Generating in short blocks with quick review costs a little more attention and saves entire reels of unusable footage.
What matters most for a cinematic result?
Light and lens. Composition gets you a competent frame; motivated light and a deliberate focal length get you a frame that feels shot rather than generated. If you only fix one thing, fix the lighting description.
How do I keep a sequence from feeling repetitive?
Keep the style block fixed but vary shot size, camera height and subject action. A sequence needs contrast in scale: wide for context, medium for behaviour, close for detail. The look stays consistent while the visual information changes.
Put the workflow into motion
The pattern behind every good AI video is unglamorous: decide the shot, specify the frame, describe one motion, review, repeat. Models will keep improving, and prompts that rely on novelty will keep expiring. Craft compounds instead.
Orelon is built for exactly this rhythm, with a still-first workflow that keeps your look consistent from the first frame to the final cut. Start with a single reference image in the AI image generator, give it one clear action in the AI video generator, and borrow pacing ideas from the video templates when you need a structure to build against. The goal is not to generate more clips. The goal is to generate the right frame, then let it move.