How to generate a cinematic fly-through of a place that does not exist

Last updated 2026-08-25 · 7 min read

What is the short version?

A fly-through of a facility you cannot film — because it is conceptual, unbuilt or simply not real — is a sequence of generated shots joined by continuous camera movement. The brief has to describe the space and the motion separately, because the generator treats them as different instructions.

1. Go to Vizard Agent and describe the space in detail.
2. Say how the camera moves through it.
3. Give the narration audio, or ask Vizard Agent to write one.

What do you need before you start?

A written description and, ideally, the narration. Vizard Agent generates every shot, sources the score and builds the captions, so no footage or renders are needed. Being explicit about what should not appear matters as much as what should, since generators add people to empty spaces by default.

• The space. What it is, what it contains, what scale it is.
• The camera movement. Travelling forward, orbiting, rising.
• What must not appear. People, branding, recognisable products.
• The narration. Yours, or Vizard Agent writes it.
• The length. Fifteen seconds per section is a workable unit.

What do you type into Vizard Agent?

Describe the space, the movement and the exclusions. Vizard Agent reads each of those as a separate constraint, and the exclusion list is the one people leave out — which is how an abstract data visualisation ends up with a figure walking through it.

Prompt

Variants worth knowing:

• Delivered in parts. Two fifteen-second sections rather than one long file.
• Your own narration audio. Vizard Agent will find its exact boundaries in the file.
• Abstract rather than literal. For anything representing data or networks.

What does Vizard Agent actually do?

Here is the order Vizard Agent worked in on a real fly-through. Two details are worth taking: it trimmed the narration to its true spoken boundaries first, and it regenerated a shot that had quietly ignored part of the brief.

1. Analyses the supplied audio and finds the exact start of the narration inside it.
2. Checks the word timestamps around that section and extracts the timing precisely.
3. Extracts the voiceover portion from the audio file and uploads it for verification.
4. Verifies the trimmed voiceover's quality and its start and end boundaries.
5. Generates the four cinematic scenes and the background score concurrently.
6. Generates the sound effects and extracts sample frames from every shot.
7. Reviews the generated scenes as a contact sheet for quality and consistency.
8. Regenerates the scene that contained a humanoid figure the brief had excluded, and checks the replacement.
9. Generates word-accurate captions, measures the voiceover, music and effects in LUFS, assembles the cut, then repeats the whole process for the second section.

Step 8 is the check that keeps a brief intact. A generator asked for an abstract network will often add a person for scale, and nobody notices until the client does, so the exclusion has to be verified rather than assumed.

What does the result look like?

From the run this page is written from, probed on one delivered section: 1080x1920, H.264, 30fps, 15.3 seconds, AAC audio. Vertical, fifteen seconds, four generated scenes moving through the space with a trimmed narration, animated captions, a score and sound design measured against the voice.

Delivering in two fifteen-second parts rather than one long file suits how this content gets used. Each section stands alone on a feed, and the pair plays as a sequence when they are posted together.

When does this not work well?

Everything on screen is generated, which is exactly right for a space that does not exist and wrong for anything that has to be accurate. Vizard Agent builds convincingly, and the gap between convincing and true is the risk here.

• This is not a visualisation of your facility. It evokes a type of space, not a specific one.
• Do not present it as a render of a real design. Architects and clients will treat it as one.
• Generators add people. Say explicitly if the space must be empty.
• Continuity across shots is approximate. Four generated scenes are a sequence, not one continuous take.
• Scale is hard to convey. Without people or objects for reference, size reads as ambiguous.

How do you fix a result that came back wrong?

Name the scene. Vizard Agent keeps every generated shot, the trimmed narration with its timings, the captions and the loudness measurements, so replacing a scene or re-timing the cut is a re-render rather than a fresh generation round.

• "There is a person in shot four." Regenerated against the exclusion, as here.
• "The camera move feels wrong." Re-described and regenerated for that scene.
• "The music covers the narration." Re-mixed against the measured stems.

How does Vizard Agent compare to doing it yourself?

By hand this means commissioning a 3D render, which prices most projects out, or generating four clips and accepting whichever came back closest. The figure standing in the middle of your abstract network is the sort of thing that survives to the client presentation, because nobody re-reads the brief once the shots look impressive.

Common questions

Can Vizard Agent make this look like our actual facility? No. It evokes a type of space rather than reproducing a specific one.

Do I need to supply the narration? Either. Vizard Agent will write one, or find the spoken section inside audio you already have.

Why trim the narration audio? Because a supplied file usually has silence, a count-in or an unrelated section around the part you want. Finding the true boundaries means the shots can be timed to speech rather than to padding.

How long should a section be? Around fifteen seconds. Two sections play better than one thirty-second file.

Will there be people in the shots? Only if you want them. Say so explicitly either way, because generators add them by default.

Can Vizard Agent do a continuous camera move? It generates each scene separately, so the sequence reads as continuous rather than being one unbroken take.

Does Vizard Agent add sound effects? Yes, measured against the voice and the score before mixing.

Can Vizard Agent match a brand's look? It can follow a described palette and style. Do not put a real logo on a space that does not exist.

Does this work for other imagined spaces? Yes. A factory, a laboratory, an unbuilt building, an interior concept — the method is the same.

Does Vizard Agent check the finished video? Yes. It measures the final loudness, reviews a contact sheet and verifies the pacing and sync for every section.