On February 27, 2026, Google released the image-generation model Nano Banana 2 and immediately made it the default in the Gemini app and AI Mode (our launch report). A continuation of the Imagen and Gemini Nano line, its headline is speed: knowledge distillation compressed the parameters to roughly a third of its predecessor, and with an MoE architecture plus TPU/Edge TPU acceleration it generates a 1024x1024 image in 2–3 seconds, with on-device inference on Android in some scenarios. Multilingual prompts (including Chinese and Japanese), content-safety filtering and watermark tracing are built in — see also the Ars Technica-sourced launch coverage.
WIRED's hands-on test (full report) scored it 8.5/10: text-to-image delivers a high-resolution result in about 5 seconds with automatic lighting and perspective correction. Editing is the highlight — swapping backgrounds or adding spacesuits, it intelligently adjusts poses and lighting for consistency; inpainting to remove passers-by, style transfer and 4K upscaling all run smoothly; face generation shows far fewer artifacts than its predecessor.
The weakness is equally clear: hallucinations remain frequent in complex scenes (a banana sprouting wings but casting no shadow), with roughly 85% accuracy under the test conditions. The verdict — "usable as a productivity tool, but outputs still need human review" — was vindicated in a different way by the Google Earth incident six months later.
On July 1, 2026, Google shipped Nano Banana 2 Lite: 0.8 seconds per image on average, about 40% faster than Nano Banana 2, 30% cheaper to run, with only 60% of the parameters — while FID and CLIP Score dropped less than 2 points (release report). Another test found Lite takes 2–3 seconds for a 512x512 image versus about 8 seconds for Nano Banana 2 and 15+ for the high-end Imagen line.
The cost is texture detail and complex lighting — Google's blind tests found Lite "good enough" in 80% of everyday scenarios. It ships via Google Cloud Vertex AI and AI Studio and comes preinstalled on new Android builds (comparison report). The selection logic is simple: Lite for high-frequency lightweight jobs like social images and UI icons; the full model when photorealism matters.
On July 31, 2026, Google wired Nano Banana 2 into Google Earth, letting users generate and overlay AI content on real satellite maps from text prompts. That same day, digital-ethics researcher Henk van Ess used it to fabricate two images — refugees massing at the Mexican border, and a bomb crater beside a Gaza hospital — unrelated to any real event yet carrying the visual authority of satellite imagery, and showed that Google's digital watermark was ignored by existing detection tools. On August 1, Google withdrew the feature — 24 hours after launch. Full account in our report.
The core problem: once generation is bound to real geographic coordinates, fabricated content acquires "geographic credibility" — far more dangerous in circulation than an ordinary deepfake.
The technical lesson of the Google Earth episode is about ordering: Google relied on after-the-fact watermarking plus topic filtering, with no geographic-sensitivity or authenticity checks at generation time — so any prompt produced convincing output, and the watermark was shown bypassable within 24 hours. The industry direction is now clear: protection must move up into the generation stage, with post-hoc detection as a supplement only.
For what watermarks can and cannot prove, see our guide: AI watermarking explained. A practical note for creators: content generated with Nano Banana 2 carries detectable marks by default — before commercial use, check your platform's and jurisdiction's AI-content labeling requirements.
It is the default image model in the Gemini app and AI Mode — regular users just type a prompt in Gemini. Developers can call the API via Google Cloud Vertex AI and AI Studio (the Lite version costs less), and Lite comes preinstalled on new Android builds. Everyday use inside Gemini needs no separate payment; the API is usage-billed.
Lite, released in July, is the lightweight build: about 60% of the full model's parameters, 0.8s per image on average (roughly 40% faster), 30% cheaper, with under 2 points of FID/CLIP degradation — at the cost of texture detail and complex lighting. Lite covers everyday images and icons; use the full model when photorealism matters.
The feature, launched July 31, 2026, let users overlay AI-generated content on real satellite maps. Researcher Henk van Ess fabricated convincing fake satellite scenes ("border refugees", "a bomb crater beside a hospital") the same day and showed the watermark could be bypassed by detection tools. Over deepfake and misinformation risks, Google withdrew it on August 1 — 24 hours after launch.
Yes. Google built in digital watermarking and content filtering for provenance and safety review. But the Google Earth episode proved the watermark can be ignored by existing detection tools, so it cannot be the only line of defense — verifying an image still requires source checking, and publishing AI content should follow platform and local labeling rules.
