Google's Gemini Live Now Generates and Edits Images While You Talk

Google just collapsed two of its most popular Gemini features into a single, continuous experience. Gemini Live, the real-time voice and camera assistant, can now generate and edit images on the fly while you are pointing your phone at the world and talking. No switching apps, no breaking the conversation to type a prompt, no pasting screenshots back and forth.

The flow is intentionally simple: open the Gemini app, tap the Live button, share your camera, and describe what you want to see. The model then produces or modifies an image inside the same session you are already speaking in. Google is pitching three early use cases for the rollout: visualizing room decor changes, working through math problems, and spinning up shareable memes.

What is actually new here

Image generation and Gemini Live have both existed for a while, but as separate surfaces. Live started as audio only and then gained camera and screen sharing, letting you talk with Gemini about whatever is in front of you. In less than a year, Gemini Live transformed from an audio only interface to a truly multimodal and dynamic conversational experience where you can discuss images, files, and YouTube videos, with continuous video sharing as the next frontier.

Meanwhile, image creation moved into the main Gemini app through Nano Banana. Nano Banana is the name for Gemini's native image generation capabilities, letting the model generate and process images conversationally with text, images, video, or a combination, so you can create, edit, and iterate on visuals with unprecedented control. What changed is that this image pipeline now runs inside the live, streaming voice channel.

How it works under the hood

The underlying recipe was previewed earlier through Google's developer docs and community examples. The pattern extends Gemini Live API's real-time conversation capabilities to image and video generation through function calling, combining Gemini Live API, Gemini 3 Pro Image, Veo, and ffmpeg so users can create and edit images, generate videos, and produce short movies through natural voice and text interaction.

In other words, Live is the conversational orchestrator. When you ask for an image, it calls out to an image model, currently the Gemini 3.1 Flash Image family marketed as Nano Banana 2. That model is optimized for speed and high-volume use cases through advanced reasoning, designed to tackle complex, multi-turn creation and modification tasks, which is what a back-and-forth voice session demands.

Where it shines, where it stumbles

The Nano Banana 2 model behind the feature has a fairly well-documented strengths profile. It brings improved world knowledge for diagrams and infographics, character consistency across generated images including blending multiple images, local edits for quick specific changes to parts of an image, better text rendering with more accurate spelling, and enhanced instruction following for complex requests.

The weak spots are equally explicit. It can still struggle with small faces, accurate spelling, and fine details, and when generating infographics or representing complex data it may misinterpret information or produce factually incorrect results. Advanced features like masked editing, major lighting changes, or blending multiple images may sometimes produce unnatural results, visual artifacts, or disjointed scenes. That matters for the math use case in particular, where a wrong digit or mislabeled axis silently breaks the answer.

Access, cost, and limits

The Gemini app itself is free, and Live's camera and screen sharing have been free on Android and iOS since the last I/O wave. Gemini Live is available for free on iOS and Android, letting you chat with the AI, use your smartphone's camera to show it things, and even screen share without spending any money. The image side has the usual tiering: preview images at high resolution, with downloads at 2K resolution on a Google AI plan or 1K without one. Paid subscribers can also use Nano Banana Pro to regenerate an image, which provides additional detail, especially for images that use text rendering or for infographics.

A few hard restrictions worth noting before you build anything around it: editing images is not available to users under 18, and camera and screen sharing in Live requires an Android device with 2 GB of RAM or more, running Android 10 and up, with comparable availability on iOS.

Why it matters

The interesting shift is less about a new model and more about the interaction loop. Until now, "show me" and "make me" lived in different modes. You could point your camera at a wall and ask what paint color would work, or you could open a chat and generate a mockup, but stitching those together meant screenshots and prompt engineering.

Folding image generation into a streaming voice session turns Gemini into something closer to a collaborator looking over your shoulder. You can hold up a notebook of equations and ask it to redraw a graph, gesture at a sofa and ask to see it in three different fabrics, or freeze a frame and turn it into a meme without ever leaving the conversation. For anyone building multimodal agents on the Gemini Live API, this is also a useful reference design for how to wire function-calling tools into a real-time audio channel.