Our Take
Google's post-event breakdown of internal tool use is a competent case study, not proof that these tools work better than alternatives or unlock new production capabilities; it's product marketing dressed as technical documentation.
Why it matters
Practitioners often ask whether AI generation tools are production-ready. Google's own conference production—speaker title cards, visual branding, interactive games, custom stickers—shows the workflow from prompt to asset on a large stage. The friction points matter more than the wins.
Do this week
Producers: audit your next event's asset pipeline (titles, graphics, music, interactive content) and map exactly where Nano Banana or Gemini would replace manual work vs. require custom infrastructure (as Google did with the sticker backend and Firebase integration).
Google used Gemini and Nano Banana across seven production domains
The I/O 2026 conference—keynotes, pre-show, physical signage, interactive booths, and speaker assets—relied on Google's own generative models at nearly every stage. The company released a technical breakdown of how.
A short film called "TPU Training Day" blended hand-drawn puppets and 3D animation with Nano Banana for stylization and Gemini Omni for final compositing. The company built a custom testing tool in Google AI Studio to ensure frame consistency before generating sequences at scale. The visual brand identity (four-color gradients, interlocking icons) emerged from iterative feedback loops: Gemini took five years of past I/O guidelines and generated candidates; those outputs fed back into Nano Banana with human critique.
The pre-show featured Jellectronica, a generative music installation partnering with the Monterey Bay Aquarium. A YOLO8 model trained in Colab tracked jellyfish movement in real time on a Coral NPU, feeding positional data into Flow Music and Lyria 3 Pro to generate bass, chords, and melody dynamically. The company also used Google Antigravity to batch-generate music stems (drums, bass, chords) for efficiency.
Infinite Scaler, a playable game, generated 3D levels from 2D prompts. Nano Banana created sprite sheets; those sheets fed back into the model to generate normal, roughness, and emission maps, which the team mapped onto cardboard box geometry in WebGL. Attendees typed prompts and got playable 3D worlds within seconds.
The Antigravity Coffee Co. booth used generative UI (A2UI protocol) in Flutter to let attendees design custom latte art. Nano Banana and Firebase handled the image generation and backend reasoning. Speaker title cards—animated assets showing each speaker in custom scenarios—used Nano Banana Pro for reference assets, Veo for motion prototyping, and Gemini Omni for detailed animation sequences. On-site sticker generation ran a web game where attendees selected two prompts (or hit "I'm feeling lucky"), and Nano Banana fused them into instant printed designs.
Custom infrastructure and human curation are non-negotiable
Google did not simply hand off tasks to Gemini and ship. Nearly every workflow required hand-built tooling. The film team built a custom consistency tool inside AI Studio before batch-generating frames. The sticker game required a bespoke backend, Android bot, and real-time printing pipeline. The coffee app relied on Firebase, Cloud Functions, and Firestore to bridge frontend UI to model inference. The jellyfish music system needed a trained YOLO8 model, Coral NPU deployment, and bespoke stem generation logic.
Human curation was the binding agent. Brand identity outputs "didn't quite hit the mark" on first pass, requiring iterative micro-experiments. Speaker title cards used "detailed text prompts" to keep outputs consistent with hand-drawn reference sheets. Puppet film charm depended on preserving "tiny, human imperfections," which the AI pipelines were designed to protect, not erase.
The company's framing—"AI tools unlock creativity and offload the mundane tasks, giving the team their best hours back"—overstates the ease. Building a jellyfish-synced music system from scratch, training object detection models, and writing custom stem generators are not mundane offloading; they are engineering. Where AI saved time was in iteration speed (storyboarding variations, exploring icon styles, generating sprite sheet candidates). The payoff was velocity, not elimination of skilled work.
Map your event's asset bottlenecks before assuming AI is the answer
If you produce conferences, live shows, or marketing events, audit three things: (1) Which assets take longest to iterate? (2) How many variations do you typically generate and discard? (3) Do you have existing infrastructure to integrate model inference into your pipeline? Google had all three in place. Most shops do not.
Speaker title cards and visual branding are obvious candidates for Nano Banana or Gemini; they benefit from rapid iteration and variation. Sticker design, icon exploration, and short film stylization also show real wins. Music generation and dynamic UI, by contrast, require custom backend work and are harder to drop into off-the-shelf tools. If your event does not already have a Firebase or equivalent backend, building one for a single conference is expensive.
Start with the friction points. Ask: which asset took the longest to produce, required the most client feedback, or had the most "almost right" rejections? That is where generative tools buy the most time. The coffee app and jellyfish system are cool; the brand identity iteration and speaker cards are the replicable wins.