Google ships Nano Banana 2 Lite, betting on speed and low cost for AI images

Google DeepMind has released Nano Banana 2 Lite, officially Gemini 3.1 Flash-Lite Image, a stripped-down version of its image model tuned for low latency and low per-image cost. The pitch is volume and iteration: generate thousands of images cheaply, edit them in near real time, and keep character consistency across outputs while retaining the model’s grounding in real-world knowledge. It exposes text-to-image, editing, and multi-image composition through a single drop-in API, with one partner claiming roughly 2.7x faster generation than the full Gemini 3.1 Flash Image at tight latency variance.

The target users are developers and creative teams building real-time or agent-driven apps — interior-design mockups, generative game worlds, slide-deck automation, and interactive learning tools — where waiting on a render breaks the workflow. Google leans on lmarena.ai Elo scores and artificialanalysis.ai latency data to position the model as a strong quality-per-dollar option against competitors, while conceding it trails the heavier Nano Banana 2 on outright fidelity.

For a technical audience, the notable details are the trade-offs and provenance handling. Google flags the usual failure modes — garbled text and spelling, small faces, fine detail, unreliable infographics and data-driven outputs, and artifacts from aggressive edits like day-to-night lighting — and warns against trusting it for factual or professional content. Every image carries SynthID, Google’s invisible watermark for identifying AI-generated media, alongside content-safety filtering and red-teaming.