Describing a scene for an AI-generated video

A video generation model films what you describe to it. What you don’t describe, it invents. For a website’s background video, three or four sentences are enough, provided you put in them what matters.

The guides published by Google for Veo, by OpenAI for Sora 2 (now archived) and by Runway for its Gen models agree on the essentials. Here is their method, applied to a video that has to carry a headline.

The place, in visible details

OpenAI advises writing your prompt the way you would brief a director of photography: concrete details the camera can see. Compare:

• vague: “a bakery”;
• precise: “a light wood counter, country loaves lined up on racks, flour floating in the light”.

Add the moment and the atmosphere: the time of day, the season, the weather, the era. Google files these details under the scene’s context. Abstract words, such as “warm” or “authentic”, help little; Runway advises translating each intention into physical elements. So describe what makes the place warm: a low lamp, a fogged-up window, weathered wood.

Keep a single place per video. Runway recommends treating each generation as a single scene: changing sets mid-shot muddles the result.

The light

Light makes half the mood. Name its source and its quality:

• “low late-afternoon light through the shop window, long shadows on the floor”;
• “cold neon lights on the ceiling, reflections on wet tiles”;
• “a single workshop lamp, the rest in half-darkness”.

The Sora 2 guide also advises setting the style right at the start, for example “1970s film stock” or “documentary footage”, and naming three to five colors. An announced palette keeps the image consistent, and helps you match the site’s colors to the video’s afterwards.

The motion: the camera, then the scene

A background video moves little and slowly, so the text stays readable on top. Describe two movements at most:

• The camera’s. Film vocabulary works: static shot, slow pan to the left, dolly in, crane shot, handheld camera. Google lists them in its Veo guide.
• The scene’s. The steam from a cup, rain on a window, a curtain lifted by the air, hands kneading in the background. OpenAI advises a single action per shot, described in simple gestures.

Write what should happen, in positive form. Runway warns that a negative prompt such as “the camera doesn’t move” can produce the opposite effect: write “static camera” instead. At Google, what to avoid goes in a separate field, the negative prompt, as a simple list: “text, logo, watermark”.

The length and the loop

The models film short shots: 4, 6 or 8 seconds for Veo on Vertex AI, 5 or 10 seconds for Gen-4 at Runway. The Sora 2 guide also observes that the model follows prompts better on short shots.

For a background that loops, choose a steady, continuous motion: a very slow dolly, rain, smoke, running water. Avoid actions that have a beginning and an end, such as a door opening or someone walking into frame: the seam would show on every loop.

A calm zone for the headline

Your headline, your tagline and your button sit on the video. Plan a place for them from the description on:

• put the subject on one side, for example “the counter fills the right third of the frame”, and keep the other side simple: a wall, a sky, a bare table, a shadowed area;
• avoid tight patterns and harsh contrasts where the text goes;
• think of the phone. Veo films in landscape (16:9) or portrait (9:16); a landscape video, cropped on a vertical screen, loses both edges. Keep the essentials near the center.

What the models get wrong

The companies admit it themselves. For Veo 3, Google writes that keeping full consistency in complex scenes, or with complex motion, remains difficult. For Gen-4.5, Runway cites effects that come before their cause, such as a door opening before anyone touches the handle, and objects that vanish from one frame to the next.

Hence a few precautions:

• No text in the video. Ask for no readable sign or written menu: your headline is written in the page, where it stays sharp, is edited in one sentence and is read by search engines. Put “text, letters, logo” among the things to avoid.
• People at a distance. A close-up face draws every glance, right under your headline, and that is where a flaw in the hands or the eyes shows most. Prefer silhouettes, people seen from behind, gestures in the background.
• People described. Google found that Veo 3 tends toward light skin when the prompt specifies nothing. If people appear, say who they are.
• A simple scene. One place, one light, two movements. The busier the scene, the more likely the model is to get lost.

Saying the scene is created

Since August 2, 2026, the European AI regulation requires deepfakes to be disclosed: generated images or videos that resemble existing people, objects, places or events, and could pass for authentic. The European Commission specifies that all three criteria must be met, and that the context and the public’s expectations count. It doesn’t settle the case of a website’s background video.

Simple caution: if the scene could pass for footage of your own place, say it was created, with a discreet note near the video. And never show, in a generated scene, a recognizable real person.

In the 29webs studio, you describe this scene, and a video generation model films your page’s background. The scene is invented from your description: it evokes your trade without showing your place. The video is then re-encoded and slimmed down, without a sound track, with a poster, and it fades away when visitors have asked for less animation: the poster stays. The details are on the video background page.

In short: a precise place, a named light, a camera movement, a single action in the scene, a few seconds that loop smoothly, and a calm side for the headline. Leave the text to the page, and the faces in the distance.