Start With the Ad, Not the Model
The fastest way to waste a week on AI video is to open a generation tool before you know what the ad has to do. Generative models are extraordinarily good at producing beautiful motion, and extraordinarily indifferent to whether that motion sells anything. The teams that get consistent results treat generation as one station on an assembly line, not as the assembly line itself.
A practical AI ad pipeline has five stages, and only one of them is "generate":
- Commercial brief — the offer, the audience, the single message, the call to action, and the constraint (duration, aspect ratio, tone).
- Script and shot list — a beat-by-beat breakdown where every shot has a job.
- Asset preparation — product photography, logos, fonts, reference frames, voice track, music bed.
- Generation and assembly — producing shots, selecting takes, cutting, sound design, captions.
- Review, delivery, and testing — approval loops, export variants, performance readout.
The reason this ordering matters is that AI generation changes the economics of stages 4 and 5 dramatically while leaving stages 1 through 3 almost untouched. You still need a reason for the viewer to care. You still need to know that the first three seconds must show the product. What changes is that a shot which once required a location permit, a crew, and a week of scheduling can now be produced in an afternoon — if you can describe it precisely.
That last clause is the whole skill. Describing a shot precisely enough for a model to render it usefully is a craft, and it is closer to writing a shot list for a cinematographer than to writing a search query.
Map Your Ad Formats to the Right Generation Approach
Different ad problems call for different generation techniques. Choosing the wrong one is the most common source of disappointment, because a technique that excels at atmosphere will fail at product fidelity, and vice versa.
Text-to-video
Best for: establishing shots, lifestyle mood, abstract backgrounds, transitions, environments that do not exist or would be expensive to shoot.
Weakness: precise product representation, legible text, and consistent characters. Never rely on text-to-video alone to show a specific SKU with accurate packaging.
Image-to-video and reference-driven generation
Best for: animating a hero product shot, bringing a still photograph to life, adding camera movement to a designed frame, giving a brand photographer's image a few seconds of motion.
This is usually the highest-leverage technique for product advertising, because the product is already rendered correctly in the input frame. The model's job shrinks to motion, lighting continuity, and camera behavior — a far easier task than inventing the object.
Video-to-video and restyling
Best for: repurposing existing footage, changing the look of a shoot you already own, adapting a horizontal commercial into a vertical cut with new framing, generating stylistic variants from one performance.
If you have usable footage, this is almost always cheaper and more brand-safe than generating from scratch.
Avatar and talking-head generation
Best for: spokesperson segments, explainer voice-over, localized versions of the same script in multiple languages, testimonial-shaped content where a real presenter is unavailable.
Watch for the uncanny threshold. A slightly stiff avatar in a 30-second ad can cost more trust than the production savings are worth. Keep avatar segments short, cut away often, and pair them with strong b-roll.
Hybrid: generated footage plus real capture
Best for: almost every serious campaign. Use generated shots for environments, inserts, and impossible camera moves; use real capture for hands, product interaction, packaging legibility, and any frame that must be legally accurate. The audience rarely notices the seam when the lighting and color match.
| Ad need | Recommended approach | Why |
|---|---|---|
| Product hero shot | Image-to-video from a studio still | Product stays accurate |
| Lifestyle atmosphere | Text-to-video | Cheap iteration, no fidelity risk |
| Multi-language presenter | Avatar plus generated b-roll | One script, many markets |
| Reformat to vertical | Video-to-video with reframing | Reuses approved footage |
| Impossible camera move | Text-to-video or video-to-video | No physical alternative |
Pre-Production: Brief, Script, and Shot List
A generation session without a shot list becomes a folder of attractive clips that do not cut together. Write the shot list first, even if it is rough.
The one-message rule
Every ad carries one primary message. If your script needs two, you have two ads. Write the message as a single sentence: "This serum visibly reduces redness in seven days." Every shot either supports that sentence or gets cut.
Beat structure for short-form ads
- 0–2s — Pattern interrupt. Movement, an unusual frame, or a direct question. This beat must work with sound off.
- 2–5s — Problem or desire. Show the situation the viewer recognizes.
- 5–12s — Product entry. The product appears clearly, in focus, lit well.
- 12–20s — Proof. Demonstration, before/after, ingredient, testimonial, or data.
- 20–30s — Call to action. One action, stated verbally and shown on screen.
Longer formats (60–90s) simply extend the proof section with additional evidence beats. The skeleton stays the same.
Shot list columns that actually help
Build the list with these columns: shot number, duration, description, camera move, lighting, on-screen text, audio, generation approach, and status. The status column is what keeps a five-person team sane — it distinguishes "prompt written" from "generated" from "approved."
Prepare assets before generating
Collect brand fonts, logo files, approved color values, product stills at the highest resolution available, any talent releases, the music bed, and the voice track if one exists. AI tools make it tempting to skip this step. Skipping it guarantees that you will rebuild the same brand look from scratch in every session.
Prompting for Shots That Survive the Edit
A generation prompt for advertising is a compressed shot description. Vague prompts produce beautiful, unusable footage. Specific prompts produce footage you can cut.
The five-part prompt skeleton
- Subject — who or what, with distinguishing detail ("a matte ceramic coffee cup, warm off-white glaze, no logo").
- Action — the single motion in the shot ("steam rising gently, cup rotating slowly on a turntable").
- Camera — framing and movement ("macro close-up, 50mm look, slow push in, shallow depth of field").
- Lighting and mood — ("soft window light from the left, warm morning tones, gentle shadows").
- Technical finish — ("clean background, vertical 9:16, no text, photorealistic").
Example, weak: "a beautiful coffee ad."
Example, usable: "Macro close-up of a matte off-white ceramic cup on a dark walnut table, steam rising in slow curls, cup rotating slowly, 50mm lens look, slow push in, warm side light from the left, photorealistic, vertical 9:16, no text, no hands."
The second prompt is not more poetic. It is more constrained, and constraints are what make generated footage editable.
Camera language that models understand
Use standard production vocabulary: dolly in, dolly out, tracking shot, crane up, handheld, static tripod, orbit, whip pan, rack focus, tilt up. Add a lens hint (macro, 35mm, 85mm) and a depth-of-field note. These words map to recognizable motion patterns and produce more predictable results than abstract adjectives like "cinematic" — which, used alone, mostly signals contrast and shallow focus.
Lighting language
- Soft and diffused: overcast, window light, bounced light, beauty-dish look.
- Hard and dramatic: single hard key, strong shadow edge, rim light, practicals in frame.
- Product-specific: "gradient backdrop," "seamless white," "dark reflective surface," "light from below to catch glass edges."
Negative constraints matter
Explicitly exclude what you do not want: no text, no watermarks, no extra fingers, no warped product labels, no additional characters. Negative constraints are not a guarantee, but they measurably reduce retries.
Generate in batches, then select
Never evaluate a single take. Generate four to eight variants of the same prompt with minor variations — one with a slower camera, one wider, one with a different background. Selection is faster than repair. Keep a naming convention so approved takes are identifiable an hour later.
Keeping Product and Brand Consistent
Consistency is where most AI ad campaigns visibly fail. A viewer may not be able to articulate why a spot feels off, but mismatched product shapes and drifting character features register as distrust.
Anchor the product with a reference frame
If the product must appear in a shot, generate from an approved still rather than describing it in words. Packaging typography, label proportions, and material finishes are almost impossible to describe reliably but trivial to preserve when they are the input.
Lock a character sheet
For any recurring human presence, write a short character sheet: approximate age, hair, wardrobe colors, and two or three distinguishing details. Reuse the same wording in every prompt that involves that person, and generate from an approved frame when possible. Small wording changes produce a different person.
Standardize a look recipe
Write down a reusable finishing recipe — color temperature, contrast level, grain amount, aspect ratio, motion intensity — and append it to every prompt. This is the equivalent of a LUT and a grade note. It is what makes shots from different sessions feel like one campaign.
Protect legal exposure
Do not generate recognizable public figures, trademarked characters, or imitations of another brand's distinctive visual identity. For regulated categories such as health, finance, and alcohol, keep claims in the on-screen text and voice-over, which you control and can review, rather than trusting a generated frame to depict a result accurately.
Editing, Sound, Captions, and Delivery Specs
Generated footage becomes an ad in the edit. Budget your time accordingly — the cut, the mix, and the captions typically take as long as the generation.
The assembly pass
Lay approved takes on a timeline in shot-list order, ignoring polish. Get a rough cut with the right rhythm before refining anything. Most AI ad cuts fail for pacing, not image quality: generated shots often look best when they are shorter than a human editor's instinct suggests, especially in the first three seconds.
Sound design
Sound carries more perceived production value than image resolution. Layer three things: a music bed, a voice-over or on-screen-text-only treatment, and spot effects (whoosh on a transition, soft tick on a text reveal, ambience under environments). Generated audio tools can produce usable ambience and effects, but a real voice recording still beats synthetic narration for most brand work.
Captions and text
Burned-in captions increase completion rates on muted autoplay feeds. Keep them high-contrast, positioned inside the safe area, and never rely on a generated frame to display accurate text — add typography in the edit where you control kerning, spelling, and legal fine print.
Export matrix
Prepare the master, then derive variants:
- 9:16 vertical, 1080×1920 — short-form feeds.
- 1:1 square, 1080×1080 — some feed placements.
- 4:5 portrait, 1080×1350 — feed-first placements with more vertical room.
- 16:9 horizontal, 1920×1080 — pre-roll, connected TV, embedded player.
When reframing, do not simply crop. Regenerate or re-frame the key product shots so the subject sits in the vertical safe area, and keep the first-frame hook inside the center square that every feed crops around.
Bitrate and length hygiene
Export at a high bitrate for the master and let the platform transcode. Deliver cutdowns at 6, 10, and 15 seconds from the same master so a single shoot feeds several placements.
Review Loops and Version Control
AI velocity creates a review problem: stakeholders can be shown fifty variants, and fifty variants produce paralysis.
Present in rounds, with a decision frame
Show three options per round, each with a stated trade-off ("Option A leads with the problem, Option B leads with the product, Option C leads with the testimonial"). Ask for a decision on structure first, polish second.
Separate structure comments from detail comments
Feedback like "the logo is small" and "the whole second half is too slow" need different owners. Route them separately or the edit will churn.
Version names that survive a week
Use a consistent pattern: campaign_platform_aspectratio_v03_approved. Store the prompt that produced each shot next to the shot. Six weeks later, when someone asks for the same look in a new market, the prompt is the reusable asset.
Approval gates
Rough cut approval, picture lock, then mix and captions. Do not allow caption edits after picture lock to trigger a re-generation cycle — captions live in the edit, not in the model.
Testing, Measuring, and Iterating
The advantage of cheap generation is that you can test creative rather than argue about it.
- Hook testing. Produce three different first-three-seconds openings for the same body and run them against each other. Hooks move performance more than any other variable.
- Format testing. Compare the vertical cut against the square cut in the same placement before assuming which wins.
- Message testing. One ad leads with price, one with convenience, one with proof. Let the data pick the angle, then invest production quality in the winner.
- Watch time as a diagnostic. A sharp drop at second four usually means the product entry is late or the transition is jarring.
- Iterate the winner, not the loser. Most teams over-invest in fixing underperformers. Scale the ad that works by producing variations of its hook and proof.
Track results against a baseline you recorded before AI entered the pipeline, otherwise every improvement will be attributed to novelty.
Common Mistakes and How to Fix Them
Generating before scripting. Symptom: a folder of clips that will not cut together. Fix: write the shot list first, then generate only what the list requires.
Over-relying on text-to-video for products. Symptom: packaging that shifts shape between shots. Fix: generate from approved product stills.
Ignoring the first two seconds. Symptom: good creative with poor three-second retention. Fix: cut the opening to a single strong image or motion event and place the product earlier than feels comfortable.
Inconsistent color across takes. Symptom: the ad looks assembled rather than shot. Fix: apply a written look recipe and grade every shot to the same reference.
No negative constraints. Symptom: repeated retries for warped text, extra limbs, or unwanted characters. Fix: explicit exclusions in every prompt.
Treating captions as an afterthought. Symptom: clipped text, misspellings, unsafe-area collisions. Fix: caption pass after picture lock, checked on a phone screen at arm's length.
Skipping sound. Symptom: footage that feels like a demo reel. Fix: music, effects, and a clear audio arc that resolves on the call to action.
Unlimited variants. Symptom: review paralysis and missed launch dates. Fix: three options per round, decision frame attached.
No prompt archive. Symptom: rebuilding the same look from scratch. Fix: store prompts with the shots they produced.
FAQ
How long does an AI-assisted video ad take to produce?
A single 15-second ad with prepared assets typically takes two to four working days from brief to delivery: half a day for script and shot list, one to two days for generation and selection, and one day for edit, mix, captions, and exports. Complex campaigns with original product photography take longer in the asset phase, not the generation phase.
Do I still need a video editor if I use AI generation?
Yes, and their role becomes more important, not less. Generation produces raw material. Deciding which four seconds of it earns a place in a 15-second ad is editorial judgment, and it is the difference between a demo reel and a commercial.
Can AI-generated ads show a real product accurately?
Reliably only when the process starts from an approved product image and the model's job is limited to motion and lighting. Describing packaging in words rarely preserves typography or label proportions well enough for a real campaign.
How many variants of a single shot should I generate?
Four to eight per prompt, with small deliberate changes. Fewer and you accept the first result, which is usually mediocre. More and selection time exceeds the time you saved by not shooting.
What is the biggest quality risk in AI video advertising?
Temporal inconsistency — the moment a viewer notices that a shape, a face, or a light source changed between shots. Mitigate it with reference frames for anything important, a consistent look recipe, and shorter shot durations that hide small continuity gaps.
Should generated ads be disclosed?
Follow the rules of the platform you are advertising on and the regulations in your markets. Where synthetic footage could mislead about a result or a real person, disclose it. Brand trust is worth more than the ambiguity.
Can the same master work across many placements?
Yes, if you plan for it. Keep the key subject inside the center square, produce cutdowns at multiple durations from one master, and export an aspect-ratio matrix rather than cropping a hero cut on the fly.
How do I keep a consistent character across sessions?
Write a short character sheet, reuse the same descriptive wording verbatim, and start from an approved frame whenever possible. Treat the wording as a locked asset, not as something to improvise each time.
The through-line in all of this is unglamorous: the models handle rendering, and you handle intent. Teams that invest in the brief, the shot list, the reference frames, and the review discipline get ads that look expensive. Teams that chase whatever the newest generation tool produces get footage. The pipeline is the product.
