LA 25 Periwinkle Tigers

Passive ads cost brands millions. Our product helps brands obtain customers through branded interactive narrative gameplay, converting those passive audiences into active players!

Comment

Inspiration

The average online ad gets skipped within 5 seconds; audiences have learned to tune out anything that doesn't ask them for something. Interactive ads buck that trend hard: roughly 3.6x better retention than passive video, because a viewer who has to make a choice can't zone out. The catch has always been cost, a branching ad traditionally means reshooting or re-rendering every path by hand, which is why almost no one actually ships them.

Reactor changes that math. Its models generate video in real time, so a "branch" isn't a reshoot; it's a different prompt fed to the same live session.

What it does

Agent Cola is an interactive thriller built as a branded ad experience for Coca-Cola. Pepsi operatives have infiltrated headquarters and stolen the vial containing the secret recipe; you play the agent in the red jacket racing across a neon-lit skyline to get it back. Every choice — chopper or parachute, fight or flee, hold your nerve or bail, is sent live to a real generative-video model, which renders the next scene itself, in character, in the moment you pick it. Choose wrong, and the recipe ends up Compromised.

How we built it

  • Storyboarded the guardrails first. Before writing a line of prompt code, we mapped the branching structure, thereby bounding the model.
  • Built assets to match. Character boards for Agent Cola and her Pepsi rival. Scored it. A backing track plus layered sound effects tied to specific story beats.

Challenges we ran into

  • Getting Helios and LingBot to reliably follow the story. Image-to-action prompting is fragile — the same phrasing that produces a clean action shot on one model produces a static or off-brief frame on the other. We ended up encoding a strict prompt order (subject → accumulated state → one new action → lighting → camera, always last) rather than trusting the model to infer it, which stabilized results across both.
  • Latency switching between models. LingBot and Helios don't just render differently, they cut differently — LingBot's reference image is read once at world start, so a mid-story switch means a real reboot, not an instant re-anchor. We covered that gap with transition masks, so the swap reads as a deliberate cut rather than a stall.

What's next for Periwinkle Tigers

  • A better harness layer for open-world exploration Moving past a fixed branching tree toward freer, more explorable scenes without losing the guardrails that keep the model on-story.
  • Stronger guardrails to support more branching storylines Scaling the story structure up without the prompt reliability problems we hit at this scope.

Built With

Share this project:

Updates