Sound-First Vertical Video: An AI Production Workflow

Vertical short-form video is a sound-first format that most AI pipelines still treat as a picture problem. A team generates a dozen striking clips, drops them onto a timeline, and discovers the real issue: nothing in the footage has rhythm, the narration has nowhere to breathe, and the captions land on top of the subject's chin. A better model will not fix that. A different order of operations will.

This guide describes a repeatable production loop for vertical video with sound: writing to a beat map, prompting for 9:16 framing that survives the crop, generating stills before motion, cutting to audio transients, and exporting files that platform compression will not turn into mush. It is written for creators and small teams who publish weekly and cannot afford to rebuild the pipeline every time a model changes.

Why the audio track decides whether the video works

Ask ten creators why a short flopped and most will describe the image. Ask ten viewers and they will describe the sound: a flat opening line, music fighting the voice, dead silence where a punch-in should have been.

Short-form feeds autoplay with sound on, so the first two seconds of audio do the job a thumbnail used to do. If the audio does not give a reason to stay, no amount of cinematic detail recovers the view. That single fact reshapes how you should build with an AI video generator.

Three consequences worth internalizing

• Pacing is an audio decision. A cut lands because a beat, a breath, or a consonant lands there. Clip length should follow the audio rather than the other way round.
• Audio consistency matters more than visual consistency. Viewers forgive a slightly different jacket between shots. They do not forgive a narrator whose timbre shifts mid-sentence.
• Sound design is a retention tool, not polish. A short whoosh into a cut and a low thud under a reveal buy more watch time than a fourth color grade pass.

What an audio spine actually is

An audio spine is a rough, listenable version of the finished piece built from voice, music, and intentional silence, with no picture at all. It takes twenty minutes and eliminates most rework, because every visual decision afterwards has a reference point. When the spine is boring, no prompt fixes it. When the spine works, even mediocre clips read as deliberate.

Build the spine in four passes

Pass 1: speak the script against a stopwatch

Conversational narration runs about 150 words per minute. A 22-second vertical video therefore holds roughly 75 to 85 words including the hook. Read the draft aloud and delete every sentence that does not advance the argument; written copy almost always arrives 30 percent too long for the runtime you planned.

Pass 2: lock the voice before touching visuals

Record or generate the narration, then treat the file as immovable. Every later decision, from cut points to shot durations to caption breaks, references it. If you regenerate the voice halfway through, your beat map is invalid and the edit starts to feel arbitrary. Locking early is the cheapest discipline in the whole workflow.

Pass 3: convert the spine into a beat map

A beat map is a timestamped outline with two columns of intent: what the ear hears and what the eye must do. A 22-second default looks like this:

Time	Audio event	Visual job
0.0-2.0s	Hook line, no music	One tight close-up, motion already in frame
2.0-6.0s	Setup line, music enters	Medium shot establishing the subject
6.0-14.0s	Three payoff lines on the beat	Cuts every 2.5-3s, escalating motion
14.0-18.0s	Music drops out under one line	Slower detail shot, room to breathe
18.0-22.0s	Music returns, closing line	Loop-friendly final frame

Write the map before prompting. When you already know shot four lasts 2.7 seconds and needs upward motion, prompting stops being aspirational and becomes specific.

Pass 4: reserve the effects slots first

Decide where the two or three spot effects go before generating anything: a cloth rustle under a turn, a click under a product reveal, footsteps under a walk. Naming them early keeps you from generating shots that offer no surface for sound. A locked-off wide of an empty room has nothing to hang a foley cue on.

Prompting vertical framing that survives the crop

Most generation models were trained on a wide-format world. Ask for a scene and you get a composition that assumes horizontal space, then the 9:16 crop amputates it.

State the frame, not just the subject

Useful phrasing: 'vertical 9:16 composition, subject centered in the upper third, minimal headroom, compressed background, negative space in the lower third.' The lower-third reservation is not aesthetic. It is where captions, handles, and interface elements live. If hands or the product sit there, they will be covered.

Skip wide establishing shots unless they carry information. In vertical, a wide shot is mostly sky and floor. Prefer medium shots and close-ups, and let the sound imply the space you never show.

Describe camera, light, then motion, in that order

'Medium close-up, vertical 9:16, woman in her thirties lit by window light from camera left, steam rising from a cup in the lower frame, slow push-in, shallow depth of field' gives a model four separate instructions: camera position, light direction, motion, composition. 'A woman in a café' gives it a mood board.

Motion that reads on a small screen

A three-inch screen swallows subtlety, but that does not mean everything should whip-pan. Motion needs a clear direction and a visible endpoint.

• Works: slow push-in, handheld drift, a subject walking toward camera, fabric or hair moving, liquid pouring, a hand entering frame.
• Fails: fast lateral pans that blur across a narrow frame, dense crowds, tiny background action nobody can resolve, orbits around static objects.

Keep individual shots between three and six seconds. Longer generations drift, morph, and invent artifacts. If a beat needs eight seconds of coverage, use two shots with a cut on the beat rather than one long take.

Generate stills first, motion second

A still is cheap to iterate and easy to judge for composition; a video generation is neither. Produce frames in an AI image generator, approve the ones that already look right as photographs, then animate only those in the AI video generator. This two-stage habit saves more time than any single prompt trick, and it keeps you from discovering a framing error after the motion pass.

Consistency is a documentation problem, not a model problem

Character drift is the most common complaint about AI video, and it usually traces back to how the prompts were written rather than what generated them.

Build a character sheet before you generate a single frame: age range, face shape, hair, wardrobe with two fixed colors, preferred lighting, and one signature prop. Paste a condensed version into every prompt in the same word order. Where the tool accepts image references, feed the same approved portrait into every shot instead of re-describing the face in language.

For products, generate one hero still and reference it everywhere. Keep camera height broadly consistent: a bottle seen at eye level in shot one and from directly overhead in shot three reads as a different object, even when the label is identical.

Then decide what you will tolerate. A slightly different collar is invisible on a phone. A different face is not. Write that tolerance down so nobody on the team re-litigates it during every edit. A quick test helps: watch the assembled cut at phone size, at arm's length, before you approve anything. Most drift that looks alarming on a desktop monitor disappears there, and the drift that survives is the drift worth regenerating.

Cut to transients, not to the timeline grid

The frame-early rule

Place each cut about half a frame before the transient rather than after it. A cut that lands just ahead of the beat feels authored; the identical cut a few frames late feels amateur. Every mainstream editor supports frame nudging, so use it instead of dragging clips with a mouse.

Duck the bed, protect the voice

Music exists to support the read. Pull the bed 12 to 18 dB under narration so the voice sits clearly on top, and let the music rise only in the gaps you deliberately left. If you can hear the bed as clearly as the voice, the voice is losing.

Loudness, headroom, and the tail

Platforms normalize loudness, so a quiet export simply gets pushed up along with its noise floor. Target a consistent integrated loudness across your catalog (around -14 LUFS is a common streaming reference), keep true peaks under -1 dBTP, and high-pass the voice near 80 to 100 Hz to remove rumble that eats headroom. End on a short tail of room tone; a hard cut to digital silence sounds broken.

Where AI audio fits and where it does not

Generated narration is fine for drafts and for formats where the voice is deliberately anonymous. Generated music works as a bed if you edit levels manually afterwards. What generated audio will not do is match the specificity of a real performance: the half-laugh, the breath before a claim. If the hook depends on personality, record it yourself and use automation for everything else.

Captions and safe zones that do not fight the frame

Captions serve muted viewers, non-native speakers, and anyone in a loud room, and they lift completion rates because the eye has something to do while the ear listens.

Practical rules: two to four words per line, sentence case rather than all caps, high-contrast text with a subtle shadow or semi-opaque plate, and nothing placed in the bottom 20 percent of the frame where platform interface sits. Keep the top 10 percent clear as well. Burn captions in the editor rather than the generator, because you will rewrite the wording three times before you are happy and a generator cannot revise a burned-in word.

One accessible habit worth building: caption the sound, not only the dialogue. A bracketed note like [music swells] or [sharp click] costs four words and tells a deaf viewer why the cut happened.

A weekly production loop you can actually sustain

1. Script to time, not to word count. Read aloud with a stopwatch and cut to the runtime.
2. Lock the voice track. No regeneration after this point unless the concept changes.
3. Draw the beat map and the shot list. One row per shot with timecode, duration, framing, motion, subject, and the audio event it supports.
4. Generate stills. Approve composition before spending time on motion.
5. Animate approved frames in three-to-six-second pieces.
6. Assemble against the audio spine. Voice first, then clips, then effects. Nudge clips, never audio.
7. Caption and export at true 1080x1920, then archive the spine file alongside the finished cut so next week starts from a template instead of a blank page.

Mistakes that quietly flatten otherwise good shorts

• Generating before scripting. You end up with clips looking for a reason to exist.
• Six or seven shots in twenty seconds. None of them registers.
• Wide shots in a vertical frame. They read as empty space.
• A static first frame. It looks like a loading screen; start mid-motion.
• Reusing the same clip twice. Viewers notice within a second and assume the rest is recycled too.
• Music at demo volume. The narration ends up competing with its own soundtrack.
• Ignoring the final frame. Design it to loop into the first, because rewatches are the cheapest reach you will ever get.
• Exporting a scaled crop. Render native 1080x1920 rather than stretching a square.

Choosing tools without overbuilding

A workable stack is smaller than most people assume: a script document, a voice source, an image generator for keyframes, a video generator for motion, an editor for assembly and captions, and a music source you have the rights to use. Everything beyond that is optional.

When evaluating an AI video tool for this workflow, judge it on four things: whether it respects vertical framing in the prompt, whether it accepts reference images for consistency, whether it holds a three-to-six-second shot without morphing, and how fast you can iterate on a still before committing. Speed of iteration beats raw output quality, because you will discard most of what you generate.

Reusable structures remove the blank-page problem faster than a new model does, so start from video templates and the prompt library before shopping for capabilities you may never use. If you are comparing platforms, match them to a specific shot type rather than a general impression; the alternatives pages and the Orelon blog are useful for that kind of side-by-side reading.

FAQ

How long should an AI-assisted vertical video be?

Start at 18 to 25 seconds. That is long enough for a hook, one idea, and a payoff, and short enough that generation artifacts stay hidden. Extend only when the script genuinely needs a second argument, and add runtime in beats rather than in seconds.

Can the sound be generated too?

Yes for narration and music beds, with a caveat: generated voice needs a consistent speaker across the whole piece, and generated music needs to sit under the voice rather than compete with it. Treat generated audio as a draft and set levels by hand afterwards.

Why does my vertical video look cropped instead of composed?

Because the model composed horizontally. Add explicit vertical framing language to every prompt, covering composition, subject placement, headroom, and negative space, and favor medium and close shots over wide ones.

How do I keep the same character across six shots?

Write a character sheet, reuse identical descriptive wording in the same order, and use image references wherever the tool supports them. Accept minor wardrobe drift; reject face drift.

Do captions really change performance?

They change comprehension, which changes retention. Two to four words per line, placed above the platform interface and synced to the voice, will outperform auto-captions that span the full width of the frame.

What is the fastest fix for an existing flat short?

Replace the audio. Re-record the first two seconds with a sharper hook and re-cut to a beat map. Most flat videos are flat because the opening line is a greeting instead of a claim.

Turn the workflow into your default

Sound-first, vertical-first production is a habit rather than a trick. Write the beat map, lock the voice, generate stills before motion, cut on the audio, then layer two or three spot effects and clean captions. Run it twice and the sequence becomes automatic, and your output stops looking like a demo of a model and starts looking like a video someone chose to watch.

When you are ready to put the loop into practice, begin with a single 22-second piece in the Orelon AI video generator: build your keyframes, animate them to the beat, and finish the sound in your editor. Cinematic ideas in motion only land when the audio gives them a reason to move.