Google DeepMind on Building Nano Banana (Gemini 2.5 Flash Image)
Oliver Wang and Nicole Brichtova from Google DeepMind join the a16z Show to discuss the making of Gemini 2.5 Flash Image, known as Nano Banana. The model emerged from merging DeepMind's earlier Imagine family of image models—known for top-tier visual quality and specialized editing—with Gemini 2.0's breakthrough in jointly generating text and images. The combination delivered conversational, iterative image editing with the visual fidelity professionals expect. The team describes the moment they realized Nano Banana had crossed a threshold: the model went viral on LMArena despite being served only a fraction of the time, and internal demos where the model could preserve a person's likeness from a single zero-shot reference image sparked widespread experimentation. They frame character consistency, customization, and multi-turn iterative editing as the core capabilities they optimized for, and predict continued improvement in long-conversation fidelity. The conversation explores the future of visual creativity—arguing AI models are tools (akin to "watercolors for Michelangelo") that let creators spend more time on intent and craft rather than tedious editing. They discuss the spectrum of interfaces (from chat-based consumers to node-based ComfyUI workflows for power users), the likely coexistence of many specialized models rather than one model to rule them all, and the 3D-vs-2D projection debate, with Wang favoring 2D projections because human cognition and interfaces are already built around them. Force multipliers highlighted include latency, factual text rendering for educational visualizations, and brand-guideline adherence at inference time. The team flags interleaved multi-image story generation as an underrated capability and identifies raising the floor of worst-case image quality—moving from "cherry-picking" to "lemon-picking" evaluations—as their next big challenge.
Original sources
Podcast The a16z Show
Episode:Google DeepMind Developers: How Nano Banana Was Made ↗
YouTube The a16z Show
Video:Google DeepMind Developers: How Nano Banana Was Made ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.