Why text-to-video moved from demo to default
For years, generative video was a party trick. You typed a surreal sentence, waited several minutes, and received four seconds of melting hands. Nobody shipped that to a client. The gap between the demo and the deliverable was enormous: no shot control, no consistency, no resolution, no sound, and no way to reproduce a result you liked.
That gap has closed enough that text-to-video now sits inside ordinary marketing production. The reason is not that the models became magical. It is that the surrounding workflow matured. Teams learned to treat generation as one step in a pipeline rather than as the whole pipeline. Scripts still get written. Shot lists still get built. Editors still cut. The change is that a growing share of the raw footage — b-roll, product inserts, abstract transitions, character vignettes, localized variants — no longer requires a camera, a location, or a crew day.
This guide is a working playbook rather than a hype summary. It covers how to choose between video models, how to prompt them like a director instead of a novelist, how to keep characters and products consistent across shots, how to plan cost and throughput, where the technology still breaks, and how to keep legal and brand teams comfortable.
What actually changed in AI video production
Generation quality crossed a usability threshold
The practical threshold for marketing is not photorealism. It is predictability at short durations. A five-second shot that matches the brief, holds its composition, and cuts cleanly into an edit is worth more than a gorgeous twenty-second clip that drifts, morphs, or invents a second character halfway through.
Modern models are good at short, specific, physically plausible motion: a hand opening a box, a camera pushing through a doorway, rain hitting a window, a person turning to look at something off-screen. They struggle with long continuous action, complex hand interaction, and anything requiring precise continuity across a cut.
Duration, resolution, and audio improved together
The useful framing is that three capabilities arrived roughly in parallel: longer usable clips, higher output resolution, and native or near-native audio. A clip that is sharp but silent still needs a sound designer. A clip with audio but soft edges still needs upscaling. When all three land in the same tool, the marginal cost of a finished shot drops sharply — and that is what makes volume production viable.
Multiple strong models now coexist
A meaningful shift is that no single model dominates every use case. Different families lead on cinematic motion, on stylized animation, on speed, on lip-sync accuracy, or on cost per second. That means model selection is now a real editorial decision, not a matter of subscribing to the obvious winner.
Choosing a video model: a decision framework
Before you generate a single frame, decide what kind of shot you are making. The categories below map to genuinely different tooling needs.
Cinematic and photoreal shots
Use these when the shot must look like it came from a camera: brand films, hero spots, luxury product sequences, realistic human performance. Prioritize motion coherence, lens behavior, and lighting consistency. Expect slower generation and higher per-second cost. Test with a demanding prompt — a slow dolly with a human subject — because that is where quality differences show up fastest.
Fast, social-first generation
For short-form vertical content, speed beats fidelity. A model that returns an acceptable nine-by-sixteen clip in under a minute lets you test ten hook variations instead of two. Visual softness is often invisible after compression anyway. Choose this tier for ad variants, trend-reactive posts, and A/B creative testing.
Stylized and animated models
If your brand voice is illustrated, painterly, anime-influenced, or deliberately lo-fi, a photoreal model is the wrong instrument. Dedicated stylized models hold a consistent art direction across shots, which is worth far more than realism for these campaigns.
Open-weight and self-hosted options
Self-hosting matters when you handle sensitive unreleased products, need predictable unit economics at high volume, or require custom fine-tuning on a brand look. The trade-off is real: you own the GPU bill, the model updates, and the failure modes. A hybrid approach works well — hosted APIs for exploration, self-hosted for the shots you generate hundreds of times.
Specialized utility models
Not every AI video task is a whole scene. Lip-sync, background replacement, upscaling, frame interpolation, rotoscoping, and object removal are separate jobs with separate tools. Building a small stack of specialized utilities is usually cheaper and better than trying to make one general model do everything.
A quick selection checklist
- Does the model support the aspect ratios and durations my placement requires?
- Can I get consistent results across repeated generations with the same seed and prompt?
- Does it accept an image reference for product or character fidelity?
- What is the realistic cost per finished second, including retries?
- Are there licensing or commercial-use terms my legal team will accept?
- What is the data retention policy for prompts and uploads?
Prompting like a director, not a novelist
Write the shot, not the story
The most common beginner mistake is writing a paragraph of narrative. Video models do not understand plot; they understand a single moment described with camera language.
Weak prompt: A woman discovers her old diary and remembers her childhood summers, feeling nostalgic and hopeful about the future.
Strong prompt: Medium close-up, woman in her thirties seated at a wooden desk, afternoon window light from camera left, she lifts a worn leather notebook into frame and holds it still, shallow depth of field, 35mm look, slow subtle push-in.
One moment. One subject. One camera move.
Use a repeatable prompt skeleton
A structure that works across most models:
- Shot size and angle — wide, medium, close-up, low angle, over-the-shoulder.
- Subject and wardrobe — specific, consistent phrasing you reuse across shots.
- Action — one verb, one direction, one duration.
- Lens and camera movement — static, dolly in, handheld, crane up, rack focus.
- Lighting and time of day — golden hour, overcast, practical neon, hard key.
- Look and texture — 35mm grain, clean digital, muted palette, high contrast.
- Constraints — what must not appear.
Keeping the phrasing of recurring elements identical across prompts is the cheapest consistency trick available. If your product is described as "matte ceramic bottle with a brushed steel cap" in shot one, use exactly that phrase in shot forty.
Negative constraints deserve their own line
Most models accept a negative description, or at least respond to explicit exclusions. Keep them boring and specific: no text overlays, no logos, no extra limbs, no camera shake, no lens flare, no people in the background. Generic negatives like "bad quality" do very little.
Keep a prompt library
After a few dozen generations you will have a set of phrases your brand responds well to. Save them as building blocks: lighting blocks, camera blocks, product blocks, wardrobe blocks. This turns prompting from an artisanal act into a composable system, which is what lets a team of three produce what used to require a team of twelve.
A repeatable workflow from brief to publish
Stage 1: Creative brief and shot list
Start with the marketing objective, not the technology. What is the single message? Where will this run — vertical feed, pre-roll, in-app, connected TV? What is the required duration? Then write a shot list where every row specifies duration, shot type, and whether it is a generation task, a stock clip, or a live-action element.
Tag each row with a difficulty rating. Anything involving sustained hand interaction, multiple characters touching each other, or complex text on screen should be flagged for a live-action or stock fallback from the beginning. Deciding this after twenty failed generations is expensive.
Stage 2: Generate in shot-sized units
Generate short. A twelve-second scene should be built from two or three generations, not one long request. Short units give you more control, cheap retries, and clean cut points when a clip drifts near the end.
Generate three to five takes per shot with varied seeds and minor prompt adjustments. Batch your work by shot type so you can compare similar outputs side by side; evaluating shots one at a time hides quality differences.
Stage 3: Select, trim, and lock picture
Do a pass focused only on motion quality — does the movement read as physically believable? Then a pass focused on optical quality — sharpness, artifacting, edge warping. Then a pass on continuity. Only after picture lock should anyone start on sound, because regenerating a shot invalidates audio work.
Trim aggressively. The first and last half-second of a generated clip is where artifacts concentrate. Cutting into the movement usually hides the seams.
Stage 4: Assemble and sound design
Edit generated clips alongside stock, motion graphics, and any captured footage. Sound is what makes AI-generated video feel intentional rather than synthetic: ambience, Foley, and a licensed music bed do more perceptual work than another round of upscaling.
Stage 5: QA and delivery
Run a structured QA checklist before export. Then export the variants the placements require, and version-name files so the next campaign can find them.
Consistency: characters, products, and brand look
Consistency is the hardest problem in AI video marketing, and it has three layers.
Character consistency comes from a reference image or a trained identity, combined with stable descriptive phrasing. Even with references, fine facial detail will shift slightly between shots. Shoot around it: use medium and wide shots for the character, reserve extreme close-ups for shots generated from the same seed.
Product consistency is more important commercially and easier to control. Generate the product itself as a consistent 3D or high-resolution still, then use image-to-video so the model animates your asset rather than reinventing it. Never let a general model hallucinate packaging details — a wrong font on a label is a brand incident, not a rendering quirk.
Brand look consistency comes from a written style guide translated into prompt blocks: palette, grain, contrast, lensing, pacing. If every editor on your team starts from the same blocks, the campaign holds together even when shots come from different models.
Cost, speed, and throughput planning
Budget per finished second, not per generation. A usable five-second shot may require eight to twelve attempts depending on difficulty. Easy shots — landscapes, slow pushes, abstract motion — succeed in one or two tries. Hard shots — hands, crowds, dialogue with accurate lip-sync — can consume ten times the attempts of an easy shot.
A practical planning model:
- Classify shots as easy, medium, or hard.
- Estimate attempts per class from your own recent runs, not from vendor demos.
- Multiply attempts by seconds and by unit cost to get a shot budget.
- Add 30 percent for rework after client feedback.
Throughput planning matters as much as cost. Generation is asynchronous, so a team can run dozens of jobs in parallel while cutting other sequences. The real bottleneck is usually review capacity, not compute: someone has to watch every take and decide. Assign a single reviewer per campaign or quality will drift.
Where AI video still fails — and how to catch it early
Hands and teeth. Fingers multiply, teeth blur, and small props deform. Keep hands out of frame, partially obscured, or attached to a simple action.
Sustained dialogue. Lip-sync is decent for short lines and degrades with length. Generate dialogue in short bursts and cut between angles.
Legible text in frame. On-screen words are unreliable. Add text in the edit with motion graphics instead.
Physics under speed. Fast throws, splashes, and collisions tend to look wrong. Slow the action down and cut on the impact.
Continuity across cuts. Wardrobe, hair, and background details shift. Solve it with shot framing, not with more attempts.
Over-smoothing. Many models produce a slightly plastic sheen. Adding grain, reducing sharpness marginally, or grading with a film emulation layer helps more than regenerating.
Build a failure log. Every team that sustains volume production ends up with a shared document of prompt shapes that reliably break, and that document saves more money than any optimization.
Governance, disclosure, and brand safety
Three questions come up in every serious deployment.
Disclosure. Many platforms and jurisdictions now require that synthetic media be labeled. Decide a house rule and apply it consistently: a persistent on-screen label, a platform-native AI disclosure toggle, or an end-card note, depending on placement and local requirements.
Rights and likeness. Never generate a recognizable real person without documented permission. For synthetic humans, keep a record of which model produced the asset and which reference images were used.
Data handling. If you upload unreleased product images or internal scripts, check where that data goes, how long it is retained, and whether it may be used to improve a model. For sensitive work, self-hosted or enterprise-tier tools with contractual data protections are usually the only acceptable option.
FAQ
How many seconds of finished video can one person produce per day?
With a mature prompt library and a fast model tier, one editor can typically produce sixty to ninety seconds of finished, reviewed footage per day for easy and medium shots. Hard shots involving people and dialogue cut that figure significantly.
Should I use one model or several?
Several. Use a cinematic model for hero shots, a fast tier for variants and testing, and specialized utilities for lip-sync, upscaling, and cleanup. Locking into one tool guarantees you will be mediocre at part of the job.
Can I avoid generating characters entirely?
Yes, and it is often the smarter route. Product-focused, hands-free, abstract, and typographic AI video is more reliable, more brand-safe, and often more effective in paid social than synthetic human performance.
How do I make AI video feel less synthetic?
Sound design, grain, deliberate pacing, and cutting into motion do more than another round of upscaling. The uncanny feeling usually comes from missing ambience and from clips held a beat too long, not from the pixels.
What should the first test project be?
Pick a low-stakes, high-volume need: ad variants for an existing campaign, localized cutdowns, or a batch of organic social clips. Measure attempts per usable shot, then use that number to plan the next, larger project.
Getting started without overcommitting
Choose one campaign, one placement, and one model category. Write a shot list with difficulty ratings. Build a ten-phrase prompt library. Generate in short units, review in batches, and log every failure. Within two weeks you will have a realistic cost-per-finished-second figure for your own team — which is far more useful than any benchmark someone published.
From there, expand deliberately. Add a second model tier when you hit a wall on quality, add self-hosting when volume makes it economical, and add specialized utilities only when a specific problem becomes repetitive. The teams that get the most from text-to-video are not the ones with the largest tool budget. They are the ones that treat generation as one well-understood stage in a disciplined production pipeline.
