Build a repeatable AI workflow for vertical shorts: hook design, multi-image fusion, keyframe consistency, motion prompts, captions, and clean export specs.
Vertical video did not win because phones are tall. It won because the feed removed the viewer's commitment. Nobody sits down to watch a short; they arrive mid-scroll, half-distracted, with a thumb already loaded. Every craft decision in a 9:16 piece flows from that single fact, and AI video tools do not change it. They only move the hard decisions earlier, into the script, the keyframes, and the prompt.
What follows is a production workflow you can repeat weekly: define the promise, build a beat sheet, generate keyframes you can defend, fuse and animate them with restraint, then assemble something that survives compression and a small screen. The emphasis is on decision criteria, because the difference between a short that reads as intentional and one that reads as a demo reel is rarely the model — it is the choices made before anything renders.
Vertical is a different grammar, not a crop
Horizontal video assumes a viewer who already agreed to watch. Vertical assumes a viewer who has not. That asymmetry rewrites framing, pacing, and sound design at the same time.
In a narrow frame, the subject owns the center column and everything else lives in the margins. Wide establishing shots read as empty space on a phone. Two-person dialogue feels cramped unless you stack faces vertically or cut between singles. Text placed near the top or bottom disappears under interface elements you do not control. The frame rewards close subjects, readable silhouettes, and movement that travels toward or away from the camera rather than side to side — a lateral pan across a room wastes most of its travel outside the visible area.
Pacing shifts too. A landscape explainer can hold one visual for eight seconds. A vertical short usually needs a visual change every two to four seconds: a cut, a camera move, a lighting shift, a new element entering frame. That rhythm is not franticness. It is giving the eye somewhere new to land before it concludes there is nothing left to see.
And vertical is watched muted as often as not. Anything essential must exist in two channels at once — audible in the mix and legible in the captions. Design for the silent viewer first, then reward the one who turns sound on. If a shot only works with audio, it is not a shot, it is a radio segment with wallpaper.
Write the promise before you generate a frame
The most common failure in AI short-form production is generating beautiful footage before deciding what the video promises. Twelve gorgeous clips with no argument produce a screensaver.
Four hooks that survive the first second
A working opening does one of four things: it asks a question the viewer wants answered, shows a result whose process is unclear, breaks an expectation, or drops the viewer inside a moment already in progress. All four create an information gap that the remaining runtime closes.
Write the hook as one sentence before touching a generator. "This is what a two-metre desk setup looks like at midnight" is a hook. "Introducing our new product line" is an announcement, and announcements rarely survive a feed built on curiosity.
Test the sentence by asking what the viewer must know to feel the gap. If the answer is "nothing," you have a hook. If the answer is "our brand history," rewrite.
The beat sheet you will actually generate against
Almost every short that holds attention reduces to three beats: setup, development, and payoff. Write them with timecodes, not as prose, because you will generate shots to those durations.
A typical 30-second structure: 0–2s hook, 2–8s context, 8–22s development containing three visual turns, 22–27s payoff, 27–30s a closing image that either loops or invites an action. The exact numbers matter less than committing to them in advance.
Then convert the beat sheet into a shot list. Each shot gets a duration, a subject, an action, a camera behavior, a lighting direction, and the transition into the next shot. This document is what you feed into prompts, and its precision determines how much re-rolling you do later. If two people on your team cannot read the shot list and picture the same video, it is not finished.
Keep spoken copy tight: roughly 150 words per minute for calm delivery, faster for energetic vertical pacing. A 30-second short rarely needs more than 70 words of narration.
Keyframes first, because stills fail cheaply
Generate still frames before generating motion. Stills are fast, easy to re-roll, and expose composition problems instantly. A frame that is confusing as an image will be worse in motion, where the viewer has less time to parse it.
Use an AI image generator to produce one keyframe per shot at 9:16, then judge each against three questions. Is the subject unmistakable at thumbnail size? Is the center column clean, with nothing important in the outer margins? Does the light separate the subject from the background, or does it merge them into one grey mass?
Reject frames that need explaining. If you have to describe what is happening before someone understands the image, the frame is not doing its job.
Judging a frame at thumbnail size
Shrink the image to 120 pixels wide and look again. Most weak frames reveal themselves here: a subject that becomes a smudge, a prop that disappears, a composition that only works at full resolution. Since the first frame often doubles as the cover image, this is not a hypothetical test — it is the test.
Character sheets and reference frames
If a person appears in more than two shots, build a character sheet before generating anything else: a single image containing the character in three or four poses or angles, under consistent light. That sheet becomes your visual anchor. When a later keyframe drifts, you compare against the sheet rather than against memory.
Products work the same way. Generate one "hero" frame of the object in a defined orientation, then hold that orientation family across the sequence. Extreme perspective changes should be deliberate reveals, not accidents of prompting.
Multi-image fusion and the craft of the transition
Fusion — blending several source images into one continuous generated sequence — is the most useful capability in vertical storytelling, and the most abused. Used well, it turns a stack of stills into a flowing scene. Used badly, it turns a clear idea into a morphing soup.
When fusing beats cutting
Fuse when the camera is meant to travel continuously: a walk through a doorway, a slow reveal around a product, a time-of-day shift in the same room. Fuse when the subject stays identical but the environment changes, because the model's job is to interpolate the world while preserving the subject.
Cut when the idea changes. A hard cut is a punctuation mark; a fusion is a comma. If two shots belong to different arguments, blending them tells the viewer they are related when they are not. The most common fusion mistake is connecting two images that share a palette but not a subject, which produces motion with no meaning.
A transition vocabulary of three moves
Limit yourself to two or three transition types across a series and use them consistently: a hard cut on motion, a short dissolve for time passing, and a whip or push for energy. Consistency is what makes a channel feel edited rather than assembled. When every transition is a different trick, the viewer notices the technique instead of the story.
Browsing video templates early can settle structural questions fast — pick a shape, then fill it with your own material instead of inventing a format from zero.
Prompting motion instead of content
Once a keyframe is approved, the image has already decided content. Your prompt's job is to describe movement, camera behavior, and atmosphere — not to re-describe the scene, which invites the model to redraw it.
The four-block prompt
Structure every motion prompt in four blocks: subject action, camera behavior, lens feel, and light or atmosphere. "The ceramicist lifts a half-thrown bowl from the wheel, hands wet, slow push in from slightly above, 50mm feel, single warm window light from camera left, dark studio behind." Each block answers a separate question, so when a render misses you know which block to adjust instead of rewriting everything.
Generate two or three motion variants per keyframe and keep the one where movement resolves cleanly. Drift, morphing, and unrequested camera moves are the usual failure modes, and they are easier to spot instantly than to fix in the edit.
Constraints remove failures that adjectives cannot
Quality words are largely noise. Constraint language removes specific problems: "single subject, no text, no logos, no extra limbs, locked-off camera." If your workflow supports negative prompts or reference-conditioned generation, use them for the failures you actually see, not for a generic wish list.
Change one variable at a time. Vary the camera and hold the light; vary the light and hold the subject. Prompting that alters five things at once can produce a great result you are unable to reproduce — which is worse than a mediocre result you can.
A prompt library is genuinely useful here, not as a source of magic words but as a record of what already worked for a given look, so a series stays coherent across sessions and collaborators.
Consistency across shots, products, and episodes
Consistency is the hardest problem in AI short-form and where most series quietly fall apart. Audiences do not need photoreal continuity, but they do need to believe they are watching the same person or object.
The locked reference paragraph
Write one paragraph describing the character — age range, hair, build, wardrobe, distinguishing feature — and paste it verbatim into every prompt. Paraphrasing introduces variation, and variation is the enemy. The same applies to a product: material, finish, key feature, and the orientation it is always shown in.
Then separate what may change from what may not. Environment, action, and time of day can vary between shots. Wardrobe, hair, and light direction should hold. If a character must change clothes, make it a story beat with a cut, not a gradual drift across four shots.
Style notes that make a series feel like a series
Before episode three, write a half-page style note: two dominant colors plus one accent, a lens character (wide and close, or compressed and observational), a caption typeface and position, and your three permitted transitions. Keep it beside your prompt drafts. When a render looks off-brand, compare it against the note instead of against your mood that afternoon.
Style notes also make batching possible. Generate keyframes for four episodes in one session while the look is loaded in your head, then animate them in a second session. Splitting the work this way reduces the drift that creeps in when you switch between thinking visually and thinking in motion.
Assembly: captions, sound, and export specs
Edit on a vertical timeline at 1080×1920. Cut on motion so transitions feel intentional rather than arbitrary. Keep the essential action inside the central safe region — assume the top and bottom strips will be covered by interface elements and captions.
Burned-in captions are safer than platform captions and travel better across re-uploads. Placement in the upper-middle third works well; keep lines short, roughly two to three words for fast cuts and five to seven for slower passages, and hold each caption long enough to read at a glance without holding so long that it lags the audio.
Sound design does more for perceived quality than resolution. A room tone bed, one impact on the hook, and music that ducks under narration will make generated footage feel considerably more expensive than it is. If you have no sound budget at all, at minimum remove dead silence and avoid music that swells into a climax the visuals never deliver.
Export at 1080×1920, 30 or 60 frames per second depending on motion, with a bitrate high enough that gradients do not band — roughly 10–16 Mbps is a reasonable working range for vertical 1080p. Keep one master with captions and one clean master without, so you can re-caption per platform without re-rendering.
A worked example: thirty seconds, one idea
Suppose you are teasing a compact desk lamp.
The beat sheet: 0–2s a dark desk, a single click, light blooms; 2–8s close-ups of the arm and the joint; 8–22s three usage shots — reading, sketching, late-night work — all sharing the same warm light; 22–27s a wide of the finished desk; 27–30s a held frame with the product name.
Generate seven keyframes at 9:16: the dark desk, the click moment, the joint macro, three usage shots, and the final wide. Lock the palette to warm amber against cool shadow. Lock the light direction to camera left so the shadows fall identically in every frame.
Animate with minimal motion: a hand entering frame, a page turning, light shifting slightly as the lamp angle changes. Keep the camera locked on the macro shots so the product stays crisp, and reserve your one moving camera for the final wide. Fuse the desk and click frames so the light appears to arrive rather than cut in.
The result is not a commercial. It is a coherent 30 seconds that reads as deliberate — which is the actual bar for short-form.
Mistakes that flatten retention
Starting with a logo. Nobody has agreed to care yet. Open with the hook and brand the payoff.
Too many ideas. A 30-second short carries one argument and one supporting detail. Anything else becomes a list nobody remembers.
Motion for its own sake. A slow drifting camera on every shot reads as a slideshow with extra steps. Alternate locked-off frames with movement so the movement means something.
Over-generating. Twenty variants of one shot is procrastination dressed as craftsmanship. Two or three, then decide.
Ignoring the mute. If the plot lives only in the voiceover, half the audience gets a sequence of pretty images with no thread. Put the argument on screen.
Chasing continuity you do not need. Audiences forgive a jacket that shifts tone slightly. They do not forgive a character who becomes a different person. Spend your consistency effort on identity, not on background details nobody is tracking.
FAQ
Do I need video generation at all, or can I animate stills? For many shorts, subtle motion on strong stills outperforms full generation. Use generated motion where the action carries meaning — a hand, a turn, a reveal — and stills with slow push-ins where it does not.
How many shots should a 30-second short have? Typically eight to twelve, with a visual change every two to four seconds. Fewer if the shots are strong and the audio carries momentum.
Why do my AI characters change between shots? Usually because the description was reworded between prompts, or because the model was asked for a large pose change without a reference. Lock the paragraph and generate a character sheet first.
When should I fuse images rather than cut? Fuse when the camera is meant to travel and the subject stays the same. Cut when the idea changes. If you cannot state in one sentence why two shots are connected, do not blend them.
What actually drives retention? The first second and the last three. Everything between buys you the right to those two moments.
How do I keep a weekly series from looking inconsistent? Batch the work: all keyframes in one session, all motion in another, using the same style note and the same locked reference paragraphs. Consistency is mostly a scheduling decision.
Can one keyframe be reused across episodes? Yes. Recurring framing is a legitimate branding device. Hold the angle and the light, vary the action and the content.
Start with a keyframe, not a keyword
Vertical short-form rewards preparation more than raw generation power. Write the promise, build the beat sheet, generate frames you can defend, then animate with restraint and cut on motion. That sequence is what separates a channel from a folder of clips.
Orelon is built for exactly that order of operations: an AI video generator for cinematic ideas in motion, paired with image generation for the keyframes that hold a sequence together. Draft your next short's opening frame, keep the prompt library open for the look you have already proven, and see how much faster the edit becomes when the difficult choices were made before the render.