SYNTHETIC SURVEILLANCE
AMBIENT.AIAI-generated CCTV footage, built to demo and stress-test behavioral security-detection models.
A series of simulated CCTV “incident sequence” reels built to demonstrate Ambient.ai’s behavioral-detection platform. A single threat actor moves through an escalating breach across five surveillance stages (reconnaissance, perimeter breach, transit, entry, and interior climax), each captured as if by a different security camera on one coherent site, and each catching a behavioral precursor before the threat reaches an occupied area. The implicit promise: with the system deployed, the final stage never happens. Every frame is AI-generated, letting us stage scenarios that would be dangerous, costly, or impossible to film with real actors and cameras, produced safely, with no one ever at risk. The pipeline combined tightly art-directed image generation, manual image editing, AI video generation, compositing, masking, and effects work.
The flagship “Planned Breach”: one threat actor across five cameras, vehicle-gate recon, a badge-reader tailgate, the secured wing, the executive suite entry, then the interior climax.
The wider camera network across the corporate campus.
“The Campus Intrusion”: street-side recon, a weak fence junction, and transit toward an occupied wing.
The surrounding campus cameras.
Galleries, exhibit storage, and after-hours interiors for cultural institutions.
Server floors and the interior infrastructure that surrounds them.
Factory floors, raw-materials storage, and loading docks.
Control rooms, access gates, and hardened perimeters for critical infrastructure.
Hospital-site cameras: entrance, pharmacy access corridors, parking structure, and restricted clinical areas.
Banking-hall, ATM, and lot environments.
Every reel is fully synthetic: no cameras, no actors, no one ever at risk. Getting footage this directed out of today's models meant treating generative AI like a film set, a tightly art-directed, multi-stage pipeline with a human craftsman in the loop at every step.
Ideation. Every scene starts in a spreadsheet, mapped by industry and location, each with multiple prompt variations to explore. Casting wide on paper is what gives the later selection real options to choose from.
Automation. A custom Python script reads the spreadsheet row by row and runs every prompt through Google's Gemini 3 Pro Image (Preview) model, batch-generating the base images for the entire matrix, turning hundreds of variations into an automated render queue instead of one-off manual generations.
Selection. The generated options are laid out as per-scene contact sheets, so framing, lighting, and the surveillance read can be compared side by side and the strongest base plate locked before any motion work begins.
Refinement. Chosen frames move into a node-based editing graph for cleanup and updates, retouching artifacts and holding visual consistency (wardrobe, environment, and camera character) across every shot in a sequence before it's taken to video.
Animation. The locked frame is taken into image-to-video, then iterated relentlessly — each branch a re-roll fixing one specific problem (arm motion, walk direction, a drifting box) until a take reads clean and gets the green light. The model gets directed like a shoot: many takes, ruthless selection.
Generations rarely came out clean. When a take nailed one element but broke another, we cut the usable parts from multiple generations, masked out artifacts and inconsistencies, and rotoscoped key elements to keep the subject reading clearly at the top of frame. A frame-by-frame finishing pass layered on top of the AI.
