02 — The intelligence layer
The engine that changes everything but you.
Manifest Engine turns a live camera stream into a live brand world. It is built on leading real-time video foundation models, orchestrated so that the guest's identity is the one thing that never drifts.
What makes it hard
Generative video is easy to demo and hard to trust.
Most systems will happily turn a guest into a different person. That is the failure the engine is designed around. The prompt stack puts identity preservation first and last, and every world published through Manifest Studio is tested against it before it reaches a mirror.
Identity lockIdentity lock
Face, proportions, expression: kept.
The engine is instructed, before anything else and after everything else, to keep the person. Wardrobe, setting and lighting are what change.
- Face and skin tone preserved. No idealised stranger in the mirror.
- Expression tracked live. A smile in front of the mirror is a smile in the world.
- Body proportions kept. The outfit fits the person, not a template.
Outfit-only modeTwo modes
Keep the store. Change the outfit.
In outfit-only mode the real background from the camera stays exactly as it is and only the garments change: a virtual try-on that respects the room. In full-world mode the setting is transformed as well. The venue chooses per mirror.
- Outfit only. Retail, fashion, sport: the guest stays in the store.
- Full world. Museums, cinema, travel: the guest steps somewhere else.
- One switch. Toggled in the settings menu on the unit.
Prompt stackOrchestration
World, moment, look, identity.
A session prompt is layered: the brand world sets tone and palette, the moment sets wardrobe and scene, the look sets the interface, and the identity lock wraps it all. Studio authors the layers; the engine assembles them per session.
- Deterministic assembly. Same moment, same result, every session.
- Hallucination guardrails. Worlds are reviewed by people before publishing.
- Vendor-independent. The orchestration layer sits above the model.
At a glance
Manifest Engine in numbers and words.
| Rendering | Continuous real-time video, not still frames |
|---|---|
| Latency | Sub-second on a standard venue connection |
| Identity | Locked: face, skin tone, age, expression, proportions |
| Modes | Outfit only (real background kept) or full world |
| Session cap | Configurable centrally; 90 seconds by default |
| Concurrency | Multiple mirrors per venue, leases per session |
| Training | Guest footage is never used to train models |
| Transparency | AI notice shown to every guest |
Questions
What venues and brands ask first.
Manifest Engine orchestrates leading real-time video foundation models and can move between them. The identity lock, the prompt stack and the operations layer are ours.
It is designed to. Glasses, headscarves, beards, wheelchairs and children (with a guardian's consent) are all handled by the identity lock rather than fought by it.
No. Stylisation applies to wardrobe, setting and lighting. The face is not a creative variable.
The stream is processed live. The recording the guest requests is kept for seven days and then deleted.