How we used Gemini to build Google I/O 2026
Key Points
- Gemini Omni + Nano Banana composed cinematic film frames
- Sprite pipeline inferred normals/roughness/emission for 2D→3D
- Antigravity + Lyria turned live motion into generative music
Summary
Google used Gemini family models and complementary generative tools across film, visual identity, immersive experiences, and attendee apps to prototype, iterate, and ship I/O 2026 fast. Teams combined human-crafted inputs (puppetry, storyboards, UI design) with model-driven pipelines to preserve intentional imperfections, ensure consistency at scale, and enable real-time interactive experiences.
Key Points
- Core models & platforms: Gemini (Omni, Canvas, API), Nano Banana, Google AI Studio, Google Antigravity, Lyria 3, Google Flow, Colab + Coral NPU, Flutter, Firebase, Cloud Functions/Firestore/Cloud Ops.
- Film pipeline: capture performance with puppetry/3D, generate stylized frames (Nano Banana), ensure pixel-consistency via a custom Google AI Studio tool, composite with Gemini Omni to retain human imperfections.
- Visual identity: seed Gemini with historical brand guidelines, iterate outputs through Nano Banana for icon/style exploration, settle on 2D-to-3D transformable assets for consistency across physical/digital channels.
- Immersive & pre-show systems: train YOLO8 in Colab and run on Coral NPU to map jellyfish motion to Lyria-generated music; Antigravity generated music stems and agentic content production.
- Game/content pipeline (Infinite Scaler): generate sprite sheets with Nano Banana via Gemini API, infer normal/roughness/emission maps for depth, map textures onto WebGL cardboard box geometry, prototype in AI Studio then scale in Antigravity.
- Attendee apps & realtime UI: use Flutter + A2UI for adaptive interfaces, Firebase as model bridge, and Gemini Enterprise Agent Platform/Antigravity for agentic coding; single codebase strategy for cross-device low-latency UX.
- Operational practices: build testable tooling for visual consistency, use ingredient/reference sheets and detailed prompts to reduce iteration, instrument pipelines for monitoring (Cloud Ops) and rapid prototyping.
Practical takeaways for engineers
- Keep human intent: use low-fidelity captures (puppetry, sketches) as canonical sources, and apply models to enhance rather than replace those inputs.
- Enforce consistency: build verification tools (pixel checks, reference sheets) before bulk generation to avoid combinatorial rework.
- 2D→3D mapping: infer normal/roughness/emission from sprites to simplify texture-to-geometry workflows for quick WebGL integration.
- Fast prototyping → targeted scale: iterate in a studio/prototyping environment, then port stable flows to agent/runtime platforms for production.
- Use managed infra (Firebase, Cloud Functions, Cloud Ops) to reduce latency and operational overhead for model-driven UIs.