How we used Gemini to build Google I/O 2026

How we used Gemini to build Google I/O 2026

备注：以下为自动提取的正文内容。

Google DeepMind blog Google Research blog Google Developers blog Google Cloud blog See all AI updates

From the jellyfish pre-show to our “TPU Training Day” film, see how Gemini helped make I/O happen this year.

Google I/O 2026 was all about how we’re making AI helpful for everyone in new ways. But we didn’t just make announcements about our innovations in AI at I/O — we used those tools to bring I/O to life, too.

It’s both a strange and exciting moment to be building anything. We're living through an incredible shift where AI tools are getting better each month, effectively rewriting the rules of what we can create.

This year, we challenged ourselves to use the same AI we were putting on stage to out-innovate, out-create and out-efficient ourselves.

We moved faster than ever and prototyped in real-time — blending human artistry with experimental technology — with no better example than the "Timmy TPU" film.

But the reward is showing how these tools unlock creativity and offload the mundane tasks, giving the team their best hours back for the parts they are uniquely suited to do. When done right, the event is amazing on its own, and, as a viewer, you stop thinking about how AI was used. That shift is the opportunity we want to share, because people keep asking, “What can you really do with AI?”

Keep reading to learn which AI tools we used — and how we prompted them — to help make I/O 2026 happen.

The AI products & models: Google AI Studio; experimental DeepMind models; Gemini Omni; Nano Banana

What we did: We created a short film starring a bunch of TPUs getting ready to do some heavy lifting for I/O 2026.

How we did it: This project started with a question: Could we make an animated film with the simplest materials — cardboard and markers — and then use AI to bring it to life? We worked with director Laurie Rowan and Nexus Studios to blend puppets, traditional animation and AI — keeping human craft and artistry right at the heart of "TPU Training Day" (also known as "Timmy TPU").

First, we captured character performances through puppetry and simple 3D animation. This gave us full control over framing and camera movement. We then used Nano Banana to generate stylized first frames from that raw footage. To keep frames consistent, we built a custom tool inside Google AI Studio. This let us test Nano Banana frames at scale, ensuring pixel-perfect matches before generating sequences.

We merged the base animation and stylized frames using Gemini Omni and other experimental models. This elevated the film to a cinematic level while retaining the original human intent. Preserving these tiny, human imperfections is what gives puppet films their charm, and our AI pipelines were designed to protect those details.

What we did: We created the visual brand identity for I/O 2026, landing on a four-color gradient with overlapping transparencies and interlocking icons.

How we did it: Our brand identity was a close collaboration between our team and AI. We started by feeding Gemini models our past brand guidelines and five years of I/O recaps. The early outputs didn't quite hit the mark, so we ran some micro-experiments. We generated new imagery and iteratively fed outputs back into Nano Banana with feedback. We also used Nano Banana to explore icon styles. Finally, we landed on flat 2D icons that dynamically transform into hyper-textured 3D icons. This created a cohesive brand expression across keynotes, physical signage and digital apps.

Here’s a prompt we used to explore icon styles with Nano Banana:

The AI products & models: Google Antigravity; Google Colab; Google CoralNPU; Google Flow Music; Lyria 3 Pro

What we did: We kicked off the pre-show with Jellectronica, a generative music experiment in partnership with the Monterey Bay Aquarium that translated moon jelly movements into sound with Lyria 3 Pro.

How we did it: We trained a YOLO8 model in Google Colab, and then ran it on Google’s Coral NPU. This tracked jellyfish movement to control the music, which was made using Google Flow Music and the Lyria API. For example, more jellies in the bass section meant louder, more energetic bass. We also vibe-coded a mass stem generator in Google Antigravity to automate the production of music stems, like bass, chords, melody and drums.

The AI products & models: Google AI Studio; Gemini API; Gemini Canvas; Google Antigravity; Lyria 3; Nano Banana