// INSIGHTS
A Personalised AI Video for Every Attendee in 60 Seconds
6 min read
We run an AI trailer studio. An attendee steps into a kiosk, is filmed for a few seconds, and receives a cinematic trailer with themselves in it in about 60 seconds. The output lands well because it is personal and because it arrives while the person is still standing there. The interesting engineering is not the model. It is the fact that a 60-second render does not mean 60 attendees an hour, and confusing those two numbers is how these activations fall over.
Capture is the bottleneck, not the GPU
Break the session into wall-clock segments. Greeting and consent takes roughly 20 seconds. Direction and capture takes 40 to 60, because people need to be told where to look and what to do, and the first take is usually unusable. Handoff and delivery instruction takes another 20. That is a two-minute occupancy of the capture pod before a single frame is rendered. One pod therefore serves about 30 people an hour regardless of how fast the inference runs.
Rendering runs in parallel and independently. If a trailer takes 55 seconds of GPU time and you have four workers, the render ceiling is around 260 videos an hour. That is far above what one capture pod can feed it. The correct configuration is therefore multiple capture pods against a shared render cluster: typically three to four pods per four-worker cluster, which balances at roughly 100 to 120 completed trailers an hour.
That balance also explains why the 60-second promise holds under load. The guest is not waiting for a queue, they are waiting for their own render, because the cluster has headroom. The moment capture pods outnumber render capacity, the queue forms invisibly and the promise breaks for everyone behind the first backlog. Size the cluster to the pods, then hold ingest if a worker drops rather than silently extending everyone's wait.
Throughput model
What the studio consists of
A controlled capture pod
Fixed lighting, a fixed camera, a plain or lit backdrop and a floor mark. Consistent input is what makes identity preservation reliable, and lighting variance is the single largest quality driver.
A GPU render cluster
Local workers on site, not a cloud round trip. Venue uplink in an Indian expo hall cannot be trusted with a video pipeline, and local inference removes the network from the critical path.
Consent recorded at capture
Explicit likeness consent is taken on screen before filming and stored against the session ID, with stated retention and a deletion path. No consent, no capture.
Original story templates
Every trailer template, score bed, title treatment and environment is authored for the activation. No third-party film IP, no copyrighted characters, no lookalike branding.
The guest journey, second by second
Consent and template selection, 20 s
The guest picks one of two or three story templates and accepts the likeness terms on a single screen. Keeping the choice to three options prevents the pod stalling on indecision.
Directed capture, 40-60 s
An operator directs three short beats: a look to camera, a turn, a reaction. Directing beats a free-form record, because the model needs specific angles and the guest needs specific instructions.
Render and hold, 60 s
The job goes to the cluster while the guest moves to a viewing area. The pod is already accepting the next person, so the render never blocks the lane.
Delivery by QR, 15 s
A session QR opens a page with the finished video ready to save or share to WhatsApp. No email typing, no app install, no waiting for a file transfer over venue WiFi.
What usually goes wrong
Queue psychology is the first problem and it is worse here than at a photobooth, because the output is desirable enough that people will wait a long time and then be angry about it. A visible line of 30 people at a 120-second pod is a one-hour wait. Publish the wait, cap the line with a physical barrier and a numbered slot list, and close the queue before the hall closes rather than promising latecomers something you cannot deliver. Lighting is the second. Identity preservation degrades quickly with mixed colour temperature, heavy backlight or a moving spot from a neighbouring stand. The pod has to control its own light, which means a shell or drape, and it means resisting the venue designer who wants an open-plan kiosk under an atrium skylight.
Moderation applies here even though the input is a face rather than a prompt. Guests bring props, gestures, third-party logos on clothing and, occasionally, other people who did not consent. Every output passes an automated check and a human glance before it is delivered, and anything going to a public screen or a social feed passes a full hold queue. Reject politely and offer a retake. The last failure mode is expectation. The output is a stylised trailer, not a photorealistic film with the guest as a lead actor. Show a real sample on a loop at the front of the queue so people know exactly what they are waiting for. An honest sample reduces both the queue and the complaints, and the people who stay are the ones who wanted the thing you actually built.
Common questions
Is it really 60 seconds?
Sixty seconds from capture-complete to a delivered video, provided the render cluster is sized to the number of capture pods. The full guest journey including consent, direction and handover is closer to two minutes, and that is the number to use for queue planning.
What happens to the footage afterwards?
Consent is recorded at capture with a stated retention period agreed with the client, usually the event window plus a short grace period. Guests can request deletion, and the signage at the kiosk states how. Nothing is used beyond the agreed scope.
Can the trailer use our brand or a film style people recognise?
It can use your brand assets, colours, product and messaging fully. It will not imitate a specific copyrighted film, character or franchise. Every template, environment and score is original work built for the activation.
Size the pods to the footfall
Give us your expected attendance and open hours and we will model pods, render capacity and the honest queue length.