Why text-to-video became a real production layer
Marketing teams have always known that video outperforms static creative. The bottleneck was never the idea — it was the pipeline. A single 30-second brand film could swallow weeks of casting, location scouting, permits, shooting, and post-production before anyone saw a first cut. Text-to-video tools collapsed that timeline. You can now describe a scene in plain language, wait less than a minute, and get a moving shot that is good enough to show a client or ship into a paid social campaign.
Three forces pushed this from novelty to standard practice. First, short-form platforms normalized vertical, high-frequency publishing. A brand is now expected to publish several variations per week, not one hero film per quarter. Second, generation models improved at the things that used to break: hands, faces, camera movement, and simple physics. Third, the workflow around the models matured — storyboards, reference images, voice tools, and editors now connect into something that resembles an actual production line rather than a collection of experiments.
For mobile-first audiences in markets like Thailand, the effect is amplified. People scroll on phones, watch with sound off, and rely on captions to carry meaning. That combination rewards volume, speed, and clarity over cinematic polish. A clean AI-generated clip with a strong hook and burned-in captions can beat an expensive shoot that takes three weeks to deliver. The practical skill is no longer "can we make a video?" but "can we make twenty coherent variations and learn which one works?"
This guide walks through the full workflow: choosing models per shot type, writing prompts that hold up, keeping characters and products consistent, assembling a repeatable pipeline, handling sound and localization, and knowing when a human crew is still the better answer.
Choosing the right model for each shot type
There is no single best model. Every generator has strengths, and mature teams treat them as a toolkit rather than a religion. The fastest way to improve output quality is to match the shot type to the model that handles it best, then stop fighting the one that doesn't.
Product and packshot shots
Product work demands clean lines, readable labels, and controlled lighting. Diffusion-based image models such as Midjourney or Flux are excellent at generating a perfect still, and image-to-video models can then add a subtle camera push or rotating turntable move. Generate the still first, fix the label in an image editor, then animate. Trying to get a crisp logo straight from text-to-video is the most common source of wasted hours.
People, presenters, and talking heads
For a human face delivering a line, lip-sync tools (HeyGen, Synthesia, or a dedicated dubbing layer) still beat raw text-to-video for reliability. If you need a cinematic, non-speaking person — walking through a market, laughing at a table — models like Kling, Veo, or Runway handle body language and skin texture convincingly. Keep the shot under five seconds; longer generations drift.
Stylized and animated scenes
For illustrative, 2D, anime, or stop-motion looks, Pika and similar creative-first tools respond well to stylized prompts. The trick is to name the medium precisely — "flat vector animation," "clay stop-motion," "ink wash" — instead of describing realism and hoping for style.
Backgrounds, plates, and B-roll
B-roll is where AI video pays for itself fastest. Plates of rain on a window, traffic at dusk, coffee being poured, a skyline timelapse — these shots are cheap to generate, nearly impossible to shoot on demand, and endlessly reusable. Build a private library sorted by mood, city, and time of day.
| Shot type | Best approach | Typical pitfall |
|---|---|---|
| Product hero | Image first, then animate | Warped labels and logos |
| Talking head | Avatar or dubbing tool | Robotic delivery, jaw drift |
| Cinematic person | Text-to-video, short clips | Face morphing in long takes |
| Stylized scene | Style-specific model | Model drifts toward realism |
| B-roll plates | Text-to-video, batch of 10 | Repetitive motion patterns |
Writing prompts that survive generation
Prompts are production documents, not search queries. The teams that get consistent results write them like a shot list entry: subject, action, environment, camera, and light. Ambiguity is what the model fills with guesses, and its guesses rarely match the brand deck.
The five-part shot prompt
- Subject — who or what, with two or three defining details: "a woman in her thirties wearing an oversized linen shirt."
- Action — one clear verb per clip. "She lifts the cup and smiles" works. "She walks, talks, and gestures" produces mush.
- Environment — location, time of day, weather, and background density. "A small Bangkok coffee shop at 8am, morning light through frosted glass."
- Camera — lens and movement: "medium shot, 35mm, slow dolly in, shallow depth of field."
- Light and mood — "soft window light, warm tones, calm and intimate."
Write that as one or two sentences, not a paragraph. Long prompts dilute attention and often cause the model to drop the most important element.
Camera language that actually reads
Generators respond well to a small, stable vocabulary: dolly in, dolly out, pan left, tilt up, handheld, static tripod, orbit, crane, tracking shot. Avoid stacking two movements in one clip ("orbit while zooming") unless you specifically want surreal results. If the model supports negative prompts, use them for artifacts: "no text overlays, no extra fingers, no warped faces."
Common prompt failure modes
- Too many subjects. Models blend faces and swap clothing between people.
- Vague actions. "Being productive" gives you nothing usable; "typing on a laptop and nodding" gives you a shot.
- Conflicting style words. "Photorealistic anime" produces an uncanny middle ground nobody asked for.
- Text requests. On-screen words are still unreliable in generated footage. Add typography in the edit instead.
- Ignoring aspect ratio. Vertical prompts sometimes behave differently from widescreen ones. Lock the ratio before you generate, not after.
Keeping characters, products, and style consistent
Consistency is the difference between a campaign and a pile of clips. A viewer may forgive imperfect realism, but they will not forgive a spokesperson whose face changes between shots.
Reference-first pipeline
Build a character sheet before you animate anything: one front-facing image, one profile, one full-body shot, plus a wardrobe and color note. Feed those references into image-to-video or character-reference features. When a model supports seed values, record the seed alongside the prompt so a shot can be regenerated after a revision. Keep a simple spreadsheet: shot ID, model, seed, prompt version, reference files, status.
Locks for wardrobe, lighting, and lens
Choose one lens family and one lighting setup per campaign and repeat those words in every prompt. "35mm, soft window light, warm neutrals" across ten clips looks far more intentional than ten individually beautiful shots that share nothing. For product shots, keep the same background color and surface for every variant.
Version control without a studio pipeline
Name files with a convention that survives a shared drive: campaign_shot03_v2_kling_seed4481.mp4. Store prompts in a shared doc, not in your head. When a client asks for "the version from last week but with the blue shirt," the difference between a ten-minute answer and a two-day answer is a prompt log.
A repeatable eight-step production workflow
A workflow beats inspiration. This sequence assumes a 30–60 second vertical video assembled from several generated clips.
Step 1 — Brief and single-minded message
Write one sentence the viewer should remember. If you cannot state it, no amount of generation will fix the video.
Step 2 — Script and shot list
Break the script into 6–12 shots of 2–5 seconds each. Assign each shot a purpose: hook, problem, product, proof, call to action. Generate a rough voiceover scratch track so you know the real timing before spending compute on clips.
Step 3 — Storyboard and keyframes
Generate still keyframes first. They are faster, cheaper, and easier to revise. Approve the look here, at the cheap stage, rather than after ten video generations.
Step 4 — Animate approved frames
Animate stills with small, deliberate camera moves. Batch generation overnight for ten variations of the same shot, then pick the best. Expect roughly one usable clip in three; plan your schedule around that ratio instead of resenting it.
Step 5 — Select and assemble
Edit in CapCut, Premiere Pro, or DaVinci Resolve. Cut on motion, not on still frames. Keep each clip slightly longer than needed so you have handles for speed ramps and transitions.
Step 6 — Voice and sound
Add voiceover, music bed, and effects. Sound lifts mediocre visuals more than any color grade.
Step 7 — Graphics, captions, and versioning
Add captions, price tags, and end cards in the editor. Export three hook variants and two calls to action to test.
Step 8 — Review, export, archive
Run the checklist, export platform-specific ratios, and archive project files with prompts and references attached.
Sound, voice, and localization
Silent autoplay means captions do the heavy lifting, but sound still separates professional work from raw output.
Voiceover options
Record a human when credibility matters — financial, health, and government messaging especially. Use synthetic voice for high-volume variants, internal drafts, and multi-language versions. Modern text-to-speech handles Thai, English, and regional accents acceptably, but always review tone: number pronunciation, English loanwords, and product names are frequent weak points.
Music and effects
License music properly. Generate simple stingers and whooshes yourself, but keep a licensed library for anything client-facing. Keep music 12–18 dB below the voice and duck it under narration.
Subtitles and language nuance
Thai audiences often watch muted, so burned-in captions outperform auto-subtitles. Keep caption lines short, use a font with clear Thai vowel and tone-mark rendering, and test on a small phone screen. For multi-language campaigns, translate the script before generating voice — not after — because sentence length changes the edit rhythm significantly.
Quality control, disclosure, and brand safety
Create a review gate that catches the failures viewers notice instantly.
The pre-publish checklist
- Faces stable across every frame? Check at full speed, then frame by frame on suspect shots.
- Hands, hair, and jewelry physically plausible?
- Logos, labels, and packaging legible and correctly spelled?
- Text in frame correct, including any Thai script?
- Audio levels consistent, no clipping?
- Captions accurate and synced?
- Aspect ratios correct per platform?
- Claims accurate and approved by legal or compliance?
Disclosure and platform rules
Several platforms require labels on realistic synthetic media, and some regions regulate AI-generated political or endorsement content strictly. Disclose when the content could be mistaken for real footage of a real person. Never generate a testimonial from someone who has not actually said it, and never put words in a real spokesperson's mouth without consent. Keep the raw prompt log for anything that might be audited.
Budget, timelines, and when to use humans instead
AI video is cheap per clip but not free per campaign. The hidden cost is iteration time and review cycles. Budget accordingly: expect the planning, prompting, editing, and approval stages to consume more hours than the generation itself.
Decision criteria
| Situation | Better choice |
|---|---|
| 20 hook variants for paid social | AI generation |
| Product close-up with exact packaging | AI still plus light animation |
| Founder or expert delivering a claim | Human camera |
| Multi-language announcement | AI with native review |
| Emotional brand film | Human crew plus AI B-roll |
| Concept testing before a shoot | AI animatic |
Timelines that hold
A single vertical video from a written brief is realistically a one-day job for one person once the workflow is familiar: half a day for script and stills, half a day for animation, edit, and sound. A campaign of five videos plus variants is roughly a working week. Promising same-day delivery on a first-ever project is how teams end up shipping their third-best clip.
Common mistakes and how to avoid them
Generating before writing. Without a script, you produce attractive clips that do not add up to a message.
Chasing realism in every shot. Stylized, graphic, and abstract visuals often convert better and are easier to keep consistent.
Skipping the keyframe stage. Fixing a still is minutes; fixing a clip is a regeneration cycle.
Using one model for everything. Learn three tools and match them to shot types.
Ignoring the first second. The hook is the entire campaign. Generate five opening shots and test them, even if the rest is locked.
No version discipline. Unnamed files and lost prompts turn a fast workflow into an archaeology project.
Skipping sound. Muted autoplay does not mean silence is fine.
FAQ
How long should each AI-generated clip be?
Two to five seconds. Shorter clips drift less, cut better, and give you flexibility in the edit. If a scene needs eight seconds, generate two clips and cut between them.
Can I use AI-generated video in paid ads?
Yes, on most major platforms, provided the content meets advertising standards, claims are substantiated, and any required synthetic-media disclosure is applied. Check each platform's current policy before launch.
Do I need design skills to start?
Basic editing literacy helps more than design training. Understanding pacing, captions, and sound mixing will improve results faster than learning prompt tricks.
How do I keep a spokesperson looking the same across videos?
Use a character reference sheet, lock wardrobe and lighting words in every prompt, keep the same lens description, and save the seed and reference files for each approved shot.
What is the biggest quality risk?
Unstable faces and hands over longer clips, plus warped text on packaging. Review at full speed first, then frame by frame on anything with a person or a logo.
When should I still hire a production crew?
When trust depends on a real person saying real words, when the product needs physical handling, or when the brand story requires authentic emotion. Use AI for volume, testing, and B-roll; use humans for credibility.
