Write image prompts that survive the jump to video: five prompt blocks, shot-type patterns, continuity anchors, negative prompts, and an iteration loop.
A prompt is a production document, not an incantation. It tells whoever renders the frame — model or crew — where the camera stands, what the light is doing, and what must not change between shots. When a generated image looks wrong, the fix is rarely a stronger model. It is almost always a clearer brief.
This guide treats image prompting as the front half of a video workflow. You will find a block structure worth reusing, worked examples with revision notes, patterns for common shot types, continuity rules that keep a character recognizable across a sequence, and an iteration loop that leaves you with knowledge instead of a folder of random files.
What an Image Prompt Really Controls in a Video Pipeline
Every image-to-video job starts from a frame the motion model did not invent. It receives your still and extrapolates: fabric sways, light shifts, the camera creeps forward. Extrapolation is confident but not thoughtful. If the first frame leaves a jawline ambiguous, the model picks an interpretation and commits to it for the entire clip. If the light has no stated direction, the render invents a soft glow that flattens everything.
That is why prompt problems surface as motion problems. A jacket that changes shade, a background that slides sideways, a face that softens by the third second — these are continuity failures baked in before animation started.
Treat the prompt as three production decisions written down at once: casting (who is in frame and in what condition), camera (where you stand, what lens, what is behind the subject), and lighting plan (source, quality, direction, falloff). Once those three are settled in text, the video stage becomes a question of pacing rather than repair.
There is an economic argument too. Rewriting a prompt costs a minute. Discovering the flaw after a five-shot sequence is animated costs the sequence. Front-loading description is the cheapest quality control available, and it is the only one that improves with practice.
The Five Blocks of a Prompt That Survives Motion
Strong prompts are assembled rather than improvised. Each block answers one question, which means that when something goes wrong you can see which line caused it.
Subject and state. Who, plus the condition they are in: posture, gaze, wardrobe, wear, what the hands are doing. "A ceramicist in her fifties, clay drying to the elbows, eyes down at the wheel" fixes hands, gaze, and surface texture in a single clause. "A woman making pottery" fixes nothing at all.
Framing and camera position. Shot size, angle, and where the subject sits inside the rectangle. Wide, medium, close, low angle, over-the-shoulder, three-quarter profile, subject left of center with negative space on the right. Camera language changes composition without changing the subject, and it pre-loads motion: a low tracking frame tells a video model something different from a locked overhead.
Light. Source, quality, direction, background treatment. A single softbox from camera left, a cool rim from behind, a background falling off to near black. Words like "beautiful lighting" or "dramatic light" hand the decision back to the render, which will choose something generic.
Lens, texture, grade. 35mm or 85mm, shallow or deep focus, fine grain or clean digital, halation, matte finish. Grade covers color behavior: teal shadows, warm highlights, desaturated midtones. Two or three of these words usually carry it. More than five turns description into noise.
One style anchor. A film stock, a photographic genre, a material reference. One anchor unifies the image. Five anchors argue with each other.
Light deserves its own line
Lighting is the most frequently skipped block and the fastest way to change a result. Write it as a short fragment, not a mood: overcast daylight from behind, grey concrete bounce, background in haze. Name the source, the quality (hard or soft), the direction relative to the subject, and what the background does.
Light direction is also the most durable continuity anchor you have. Faces can change expression, wardrobe can change color, but if the key light keeps coming from the same side across a sequence, viewers read the shots as one scene. Break that direction only when a cut motivates it — moving from street to interior, for example.
Style anchors: one is enough, two is a fight
Stacking references feels like specifying; in practice it produces mush. Photorealistic, anime, oil painting, cinematic, hyper-detailed gives the render conflicting surface instructions, and the compromise is usually a plastic middle ground with no texture at all.
Pick the dominant anchor — photographic realism, 1970s color negative, clay stop-motion — and describe deviations inside it with ordinary language. A photographic frame with slightly desaturated color and visible grain reads as a deliberate look. The same intent written as five genre labels reads as indecision.
A Worked Example: From Vague Sentence to Shootable Frame
Here is a prompt people actually write: "A futuristic soldier standing in a ruined city, cinematic, highly detailed, dramatic lighting, 8k."
Nothing in that sentence is wrong, and nothing in it is decided. Where is the camera? What time of day? Is the armor matte, brushed, or reflective? Which way does light fall? "Cinematic" is a compliment, not an instruction, and the render cannot act on praise.
The layered rewrite:
A soldier in her late thirties, matte black exosuit scuffed across the shoulder plates, helmet carried under one arm, sweat and road dust on her forehead. Medium close-up, three-quarter angle, subject left of center with negative space on the right. Overcast late-afternoon light from behind, cool grey bounce from a concrete wall, street behind her blurred into haze. 65mm lens, shallow depth of field, fine grain, muted teal-and-amber grade. Near-future realism, photographic.
The two versions are similar in length. The second carries perhaps five times the information, and every clause is something the render can execute.
Revision notes: what changed and why
- Added age and a specific wear detail. Scuffed shoulder plates give the model a texture target instead of a generic sci-fi surface, and they will catch light consistently when the shot moves.
- Named the light direction. Backlight plus a bounce explains the rim on the shoulders and separates the subject from the background — the single most reliable way to make a frame read as photographed.
- Chose a lens instead of a resolution word. "8k" does not change composition. Focal length and depth of field do.
- Kept the background soft and non-specific. A blurred street costs the model less than a rendered crowd, and it will not slide or duplicate during motion.
- Reduced the style anchor to one. Near-future realism stays; the rest of the adjectives were doing contradictory work.
Then comes the motion clause for the video stage: slow dolly in, dust drifting through the light, helmet strap moving slightly. Physical verbs, not emotional ones. "Slow dolly in" is executable. "Epic and emotional" is not.
Prompts by Shot Type
Different shots need different weighting. These patterns are starting points to edit, not templates to copy.
Character close-up
Prioritize eyes, skin, and micro-expression. Add one imperfection — a small scar, a loose strand of hair, uneven stubble, dry lips — because flawless faces read as generated while specific faces read as photographed. State where the catch light sits, so the eyes do not go dead once the clip animates. If the character speaks or turns, describe the neck and shoulders; ambiguity there is what produces the dreaded melting jawline.
Product hero frame
Prioritize material accuracy and controlled reflections. Say whether the surface is matte, glossy, brushed, or translucent, name the single key light and its falloff, and state what the product reflects — a soft white card, a dark studio, a window. Add one subtle imperfection so the object does not look like a 3D mockup. If the product has text or a logo, describe its placement rather than assuming the render will invent clean typography.
Establishing wide shot
Prioritize scale cues and atmosphere: a foreground element for depth, a midground structure, background haze. Naming a distance relationship beats naming a mood — a figure at the far end of the corridor, roughly one-twentieth of the frame height gives the render a measurable instruction. Wide frames are also where aspect ratio matters most; decide the ratio before writing, or you will compose for one crop and deliver another.
Texture and transition inserts
These frames exist to be cut against: light through smoke, rain on glass, fabric in wind, a hand brushing past the lens. Ask for abstracted surfaces, keep the palette matched to the surrounding shots, and expect to use them in half-second bursts. They hide cuts, cover continuity gaps, and give an edit rhythm without requiring new sets.
Two-person dialogue frame
State who is in focus, who sits in soft foreground, and how the light divides them. Ambiguity in a two-face frame is expensive, because the motion model may lose track of which features belong to whom and swap them mid-clip. A simple fix: give each person one distinguishing feature and mention it explicitly in the prompt.
Negative Prompts: Cleanup, Not Construction
Negative prompts remove tendencies, not ideas. They are excellent at chronic annoyances: extra fingers, warped text, watermark artifacts, plastic skin, blown skies, duplicated limbs. They are poor at teaching structure. If the costume is wrong, no incorrect costumes will not help; a clearer positive description will.
Four rules keep them useful:
- Keep the list short — five to eight items at most, and only ones you have actually seen in your output.
- Never negate something you also want, vaguely. No darkness in a night scene creates a fight the render cannot resolve.
- Change one negative at a time, so you know which one did the work.
- Retire a negative once the positive prompt stops triggering the problem. Stale negatives suppress detail long after they stop helping.
One diagnostic worth remembering: if your output looks grey, flat, and low-detail, cut the negative list in half before blaming the model. Over-policing is a common cause of dull frames, and it is invisible in the prompt text once you have stopped reading it critically.
Continuity Across Shots: Lock Fewer Things, Lock Them Properly
A sequence needs anchors that do not change. Trying to freeze everything produces stiff, over-specified prompts that lose the small variations that make images feel alive. In practice three anchors carry most of the continuity: identity (face shape, hair, one distinguishing feature), wardrobe or material, and light direction. Palette and lens are useful secondary locks.
Write a continuity sheet
Before generating anything, write one short page containing the reusable lines for each anchor, a fixed palette description, a lens, a grade, and the aspect ratio. Every subsequent prompt becomes an assembly job rather than a fresh invention. Teams that adopt this cut their re-roll rate sharply, because differences between shots become intentional choices rather than accidents.
Add a short "do not change" list to the same page. It sounds trivial and it prevents the most common regression: a later prompt quietly rewrites the wardrobe because the writer got bored.
A pre-flight checklist before approving a sequence
- Face shape and hairline read identically at full zoom, not just at thumbnail size.
- Wardrobe values match in brightness, not only in hue name. Olive green can drift to sage between prompts.
- Light direction is consistent unless a cut motivates the change.
- Background geometry has not mirrored or slid between frames.
- Grain and grade sit at the same intensity, so cuts do not flash.
- Hands have the same number of fingers and the same level of detail in every frame.
Animate only frames you would publish
If the still is not good enough to stand alone, motion will not rescue it. It will animate the weakness and make it harder to ignore. Approve the frame first, then carry it into motion, then judge the clip on pacing rather than on description you failed to nail earlier.
An Iteration Loop That Teaches You Something
Random re-rolling feels productive. It is not, because nothing accumulates. A controlled loop costs the same time and leaves you with transferable knowledge.
- Round one: generate four variations of the layered prompt, changing nothing between them. You are measuring the model's natural spread.
- Round two: take the closest result and change exactly one line — usually light direction or framing, because those move results furthest.
- Round three: hold the winner fixed and refine only texture and grade.
- Round four: duplicate the winning prompt and swap the subject block to test the next shot against the same light and lens.
Two habits make the loop work. First, name files with the variable you changed, so a week later you still know which experiment won. Second, keep a prompt journal: one line on what you tried, one line on what happened. Ten entries in, you will start predicting outcomes before you generate, which is the real skill.
Set a stop rule as well. If three consecutive rounds on the same variable fail, the problem is the concept or the composition, not the wording. Rewrite the shot idea instead of the adjectives.
Mistakes That Show Up on Screen
| Mistake | What it looks like | Cheap fix |
|---|---|---|
| No light direction | Flat, sourceless illumination | Name source, quality, direction |
| Conflicting style anchors | Muddy rendering, no texture | Keep one anchor, delete the rest |
| Adjective stacking | Generic, glossy, forgettable frames | Replace praise with description |
| Vague subject state | Unresolved hands and gaze | Add posture, expression, action |
| Aspect ratio decided late | Cropped heads, dead space | Choose ratio before writing |
| Bloated negatives | Dull color, suppressed detail | Trim to five items |
| No motion clause | Clip drifts without intent | Add one physical verb |
The common root of most rows is the same: words that express approval instead of observation. A model cannot act on stunning. It can act on backlit, rim light along the shoulder, background falling to black.
A second, subtler mistake is writing for the still and forgetting the edit. A beautiful frame that shares no palette with its neighbors will cut badly. When you compose a prompt, glance at the two shots either side of it and match one element — grade, lens, or light quality.
Where Prompt Work Fits in the Toolchain
Prompt writing is not a separate craft from production; it is the first stage of it. A sensible split: draft and refine stills in an AI image generator, approve the frame, then carry it into an AI video generator for motion, pacing, and camera movement. Keep successful blocks in a shared prompt library so they are reusable rather than retyped, and start from a template when you need a shot shape rather than a subject.
Two practical notes about moving between stages. First, decide aspect ratio and duration before you write a single descriptive word; both constrain composition more than any adjective. Second, if you switch engines mid-sequence, re-test your locked first frame, because motion behavior differs between tools and continuity assumptions do not always transfer. Notes on those differences sit alongside the alternatives overview, and the Orelon blog covers narrower workflow questions as they come up.
FAQ
How long should an image prompt be?
Long enough to cover subject, framing, light, lens, and one style anchor — usually 40 to 90 words. Beyond that, most additions either repeat something already said or contradict it. If a prompt runs to 150 words, read it back and delete every line that describes how you feel about the image rather than what is in it.
Do negative prompts matter if the positive prompt is strong?
Less than many guides suggest. Negatives are cleanup, not construction. Spend your effort on the positive description first, then add two or three negatives for artifacts that keep appearing across several generations.
How do I keep a character consistent across many frames?
Reuse the exact same identity block — face shape, hair, one distinguishing feature — and keep wardrobe and light direction fixed. Change only framing, action, and setting between prompts. Save the block somewhere you will find it, and copy it verbatim rather than retyping from memory.
Should prompts differ for video versus stills?
Slightly. Video prompts benefit from a short motion clause: what moves, how fast, in which direction. Keep it physical — slow dolly forward, coat moving in wind, dust drifting — rather than emotional, and limit yourself to two movements so the clip does not fight itself.
Is it worth saving prompts?
Always. A saved block is a reusable asset. One saved with a note about what it produced is worth more than the image it generated, because it lets you repeat a result deliberately instead of hoping for it.
What if my subject keeps changing between shots?
Nine times out of ten, the wardrobe or identity line was rewritten rather than copied. Compare the two prompts side by side before regenerating anything; the difference is usually one word you did not notice changing.
Start With the Frame You Would Actually Publish
Prompt quality is not about vocabulary. It is about deciding what the frame contains before you ask for it: who is there, where the camera sits, where the light comes from, and which single reference holds the look together. Write those decisions down, reuse the blocks that work, and change one variable at a time until you can predict the result.
When you are ready to put that discipline into motion, open Orelon, build the keyframe in the image workspace, and carry it straight into the video canvas. Cinematic ideas move faster when the brief behind them is already sharp.

