AI video direction: why a prompt won’t make it look like cinema

A well-written prompt gets you an image. A film is something else: a chain of decisions about what to show, for how long, from where and why. AI video direction is exactly that work, applied to models that generate moving images rather than to a camera.

This piece covers what the data says about AI-made ads, the concrete levers that separate a clip from a scene, a shot-by-shot before-and-after and a checklist for before the first frame.

The problem isn’t AI, it’s the lack of direction

Merriam-Webster picked “slop” as its 2025 Word of the Year: “digital content of low quality that is produced usually in quantity by means of artificial intelligence”. The word captures a widespread irritation, and advertising is not immune.

Back in December 2024, a NielsenIQ study of more than 2,000 people, around 150 of them measured with EEG, found that viewers intuitively spotted most AI-generated ads and described them as “annoying”, “boring” and “confusing”. Even the ads rated high quality triggered weaker memory activation than traditional ones.

In January 2026, IAB measured the gap between the people who make ads and the people who watch them in the US: 82% of ad executives believe Gen Z and Millennials feel positive about AI-generated ads, yet only 45% of those consumers actually do. Among Gen Z, 30% describe brands that use them as “inauthentic”.

The most useful figure comes from DoubleVerify, published at the end of July 2026 and based on 22,000 consumers in 22 markets. Of those, 42% say low-quality or “uncanny” AI ads would hurt their opinion of a brand, while 40% view polished, professional ones positively. As CEO Mark Zagorski put it: “AI itself is not what determines engagement. The quality of the output does.” In video, quality has a precise name: direction.

What AI video direction means in practice

Direction isn’t an adjective you bolt onto a prompt, like “cinematic”, “epic” or “4K”. It’s a series of decisions made before, during and after generation. These are the first to lock.

One intention, one idea

Every video has to carry a thesis you can state in one sentence. If you don’t know what the viewer should feel at second 3 and at second 20, the model won’t know either: it will fill the gap with the most probable image, which is the one everyone has already seen.

Shot list and camera language

Break the scene into shots: framing, angle, lens, movement, duration. A locked-off 35mm describes a place; an 85mm on a slow push-in describes a thought. Even Google Cloud’s prompting guide for Veo 3.1 (October 2025) puts cinematography first in its prompt structure, ahead of subject and action.

Light and palette

Decide where the light comes from, its colour temperature and how it changes across the piece. A palette fixed in advance makes ten clips generated at different moments feel like one film.

Performance and timing

Models tend to make everything move, all the time. Direction also works by subtraction: a look held one second longer, a beat before the line, one gesture instead of five.

Continuity

Face, wardrobe, props and locations must stay identical from one shot to the next. That’s why they’re defined up front, with character sheets and reference images.

The film is made afterwards: takes, edit, sound, colour

Generating is like shooting: you produce footage, not the film. PJ Accetturo, who made an ad for Kalshi with Veo 3 that ran during the 2025 NBA Finals, told NPR that it “took about 300–400 generations to get 15 usable clips”. He added: “Just because this was cheap doesn’t mean anyone can do it.”

The same goes for hybrid work: for a live-action Under Armour spot, Chilean production company Trópico created 530 assets in Runway during post. Their description of the process: “You’re directing instead of just executing.”

• Take selection: drop anything that gives the medium away, such as warping hands or empty eyes, however good the framing.
• Editing: rhythm is decided in the cut. Cut on the gesture, not where the generated clip happens to end.
• Sound: ambience, silence and music give images weight. Sound is often what makes a scene believable.
• Colour: the final grade pulls clips from different engines and prompts into a single look.

Before and after: the same scene, with and without direction

Take a simple brief: a 12-second spot for a neighbourhood café. Here’s the prompt-only version.

Cinematic shot of a woman drinking coffee in a beautiful café at sunset, ultra realistic, 4K, epic lighting.

The result is predictable: a model with no story, golden light everywhere, a random camera move, no cut. A pleasant image that could belong to any brand, which is exactly the problem.

Now the directed version. The idea in one sentence: a nurse’s first break after a night shift. Four shots, and a palette that moves from the blue of dawn to the warmth of first light.

1. Wide shot, 35mm, locked off, 4 seconds. The café has just opened, chairs still on the tables. She walks in, in uniform, alone. Sound: the espresso machine warming up.
2. Macro insert, 2 seconds. Hands around the cup, steam rising. Silence.
3. Close-up, 85mm, slow push-in, 4 seconds. She closes her eyes and stays there a second longer than she needs to.
4. Reverse through the window, 2 seconds. The city wakes up, the first warm light reaches her face. Logo and a single note of music.

Face and uniform are locked with a character sheet. Each shot gets several takes, and only those with clean hands and living eyes survive. The edit cuts on the gesture; the grade ties it together. The model is the same. What changes is who decides.

Tools give you controls, not choices

Tools increasingly speak the language of the set. Flow, launched by Google in May 2025, offers direct control over camera motion, angles and perspectives. Higgsfield’s Cinema Studio 4.0 (Higgsfield is one of our partners) provides more than 30 camera movement presets and over 50 colour palettes. But a menu of 50 palettes won’t tell you which one serves your story.

People working at the top end start before the prompt. For the short film “ANCESTRA”, which premiered at Tribeca in 2025, Google DeepMind describes a team of more than 200 people. Director Eliza McNitt wrote the script and worked with a storyboard artist; photos her father took on the day she was born, described by Gemini, became the basis of the look. The generated newborn was placed into live-action footage, then finished with traditional VFX and colour grading.

Obsidian, a studio working for Disney and Nike, is blunter still in an interview published by Runway in January 2026: “we don’t start our projects with a prompt. We start with hands, with a pencil.” Their first hires were storyboard artists from Hollywood.

Before the first frame

Whether you’re producing or commissioning an AI video, ask these questions before anyone mentions models:

• What’s the idea, in one sentence?
• Is there a shot list with lenses, moves and durations?
• Are look, light and palette locked in a moodboard?
• Are faces and locations defined once and for all?
• Who selects the takes, and by what standard?

KASNO’s video “Is AI killing cinema?” has the line that sums it up: AI doesn’t create artists, it shows who they are. It’s the principle behind “Powered by AI. Directed by humans.”: we work with a three-act method in which storyboard and look are approved before a single frame moves, and takes are chosen one by one in production. To see how that translates into ads, brand films or reels, start with our services.

Frequently asked questions

Can you make a cinematic AI video without a director?

You can get individual beautiful images, but a sequence that holds together needs someone deciding the idea, the shots, continuity and the edit. The job title may vary; the work is still directing.

How much does the prompt matter in AI video?

About as much as a precise note to a cinematographer: it’s necessary, but it comes after the idea, the storyboard and the look. A good prompt executes a decision well; it doesn’t replace it.

Can audiences tell when a video is AI-generated?

Not always: in DoubleVerify’s July 2026 study, 56% of consumers said they cannot consistently identify AI-generated content. Flaws, however, do get noticed, and they are what shapes opinion of the brand.

Sources

1. 2025 Word of the Year: slop · Merriam-Webster, 2025-12-15
2. NIQ research uncovers hidden consumer attitudes toward AI-generated ads · NielsenIQ, 2024-12-12
3. The AI Ad Gap Widens · IAB, 2026-01-15
4. Global Study: Poor-Quality AI Content Puts Brand Trust at Risk · DoubleVerify, 2026-07-29
5. The ultimate prompting guide for Veo 3.1 · Google Cloud Blog, 2025-10-16
6. An AI video ad is making a splash. Is it the future of advertising? · NPR, 2025-06-23
7. Behind the Scenes of an Under Armour TV Commercial, Powered by Runway · Runway, 2025-10-14
8. Meet Flow: AI-powered filmmaking with Veo 3 · Google, 2025-05-20
9. Cinema Studio 4.0 — Direct Every Detail · Higgsfield, 2026-10
10. Behind “ANCESTRA”: combining Veo with live-action filmmaking · Google DeepMind, 2025-06-13
11. How Obsidian Uses Runway to Power Campaigns for Disney, Nike and Luxury Brands · Runway, 2026-01-06