How we made a five-minute brand film for $0.36

There is a five-minute film on our homepage. It has a narrator, a three-movement score, sixty-six sound effects, and fourteen scenes of animated interface and typography. The total bill for generated media was thirty-six US cents. This note explains the machinery, partly because it is the same machinery we sell, and mostly because the lessons transfer to anyone automating serious work with AI.

The genre decision came first

A brand film should be made of the brand: the mark, the type, the palette, the motion rules. Generative video is impressive, but it hallucinates detail, and a brand cannot afford hallucinated detail. So we drew a line that held for the whole production: zero generative video. Every frame is our own design system, drawn by our own renderer. The only generated media is audio: the narration and the score.

One function draws every frame

The film is a web page with a single exposed function: give it a time in seconds and it draws that exact frame. Nothing animates on its own. A renderer steps time forward a thirtieth of a second at a time, screenshots each state, and hands 9,180 frames to an encoder.

Determinism sounds academic until you use it. It means any frame can be inspected before rendering anything. It means a two-scene fix re-renders in minutes instead of hours, because frames are pure functions of time. And it means the film re-renders identically, which turns a creative artefact into something closer to a build product. The same renderer, not a similar one, prints our proposals and invoices.

The timeline is measured, not guessed

Voice comes first. Each of the twelve narrated sections is generated as its own take, then measured. Scene boundaries derive from those measured durations, laid end to end and snapped to a musical bar grid at 120 beats per minute, so every cut lands on a bar line. The picture, the score's ducking envelope and the sound-effects cue table all read the same timeline file, which makes it structurally impossible for sound and picture to disagree about time.

The score is three generated movements crossfaded under scene transitions, with the ducking baked into the audio file itself so only one layer ever touches volume. The effects are a cached library, paid for once, reused forever.

What it cost, measured

• Narration, twelve sections of text-to-speech: about $0.12
• Score, three generated movements: $0.24
• Effects, renderer, everything else: $0.00, already owned

The final master measures −14.2 LUFS integrated loudness with true peak at −1.1 dB on the encoded file. Those numbers are measured, not assumed, because the file that ships is the file that gets measured.

The process caught its own mistakes

Honesty is the point of this note, so here are the defects the pipeline caught before anyone else could:

• An independent quality check caught hot audio. The encoder adds inter-sample peaks, so a master that measured fine as a raw file exceeded spec once encoded. The fix: master with headroom for the codec, then measure the encoded file, which is the one that ships.
• An automated review pass caught small type. Text sized for the 1080p master thinned out at the 720p delivery encode. Rows were resized for the delivery resolution, not the source.
• A screenshot pass caught an invisible chart. One scene's bars had a zero-height container: the animation ran perfectly on nothing. Three model-based review passes read the empty scene as minimalism. A single per-scene screenshot pass caught it in seconds.

The lesson generalises: automated judges are useful and fallible, mechanical checks are narrow and incorruptible, and serious systems use both.

Why this matters beyond a marketing film

Swap "film" for any deliverable a business produces on repeat and the architecture holds: measure the inputs, derive the plan from measurements, make every step reproducible, check the output with independent eyes, and publish the numbers with their sources. That is how our engines are built, and the film is simply the most watchable proof of it.