AI Overviews Image Generation Lands as Google Revamps Images

What Google announced

Google marked the 25th anniversary of Google Images by redesigning the Images home page. It also added a way for people to create new pictures from scratch inside AI Overviews, without leaving the search page.21 The idea is easy to follow. If an ordinary image search doesn't turn up the picture someone has in mind, the query can serve as a prompt instead. Users can also ask for a picture directly with phrases like "help me visualize" or "create a visual."21

Google's Nano Banana image model powers the feature and turns plain-language descriptions into finished images.21 Google said image generation in AI Overviews would roll out over several weeks, in English, in every region where AI Mode already supports image creation.21 That detail matters. The feature's reach is set by AI Mode's image-creation footprint, not by the much larger area where AI Overviews appear.27

The second part of the announcement was a new Images gallery. It pulls in pictures from across the web in real time based on a user's interests. Saved collections show up as tabs above the main feed, which makes it easy to pick up where you left off.21 MediaPost compared this to Pinterest. The gallery is rolling out on desktop in the U.S., in English.21

Which model, and when

Reports don't fully agree on which model the feature uses. MediaPost described it only as Google's new Nano Banana model.21 A tracker of AI Overviews changes says 9to5Google identified it as Nano Banana 2 Lite, the fast, low-cost version of the family.27 The tracker dates Google's announcement to July 14, 2026.27

The model question got more complicated in October. On October 6, Google released Nano Banana 2.1 and put it into AI Mode in Search the same day. Search VP of Product Robby Stein told users to tap a banana icon under the search box in the Google app.110 Search Engine Roundtable's Barry Schwartz noted that Nano Banana first reached Google Lens about a year earlier. It came to AI Mode the following December and to AI Overviews in July. Schwartz didn't think AI Overviews had moved to version 2.1 yet.1 Another outlet gave users the same advice: AI Mode should produce 2.1 output, but AI Overviews may still be running the older model.2

The gap may sound minor, but it shows how Google is setting up Search. AI Mode, the conversational tab, gets the newest models on launch day. AI Overviews, which reaches far more people, appears to get them later and more cautiously. Since I/O 2026, the two have worked as one AI search experience, and AI Overviews now reaches more than 2.5 billion users a month in over 200 countries.27 At that scale, holding back makes sense.

A rough start with publishers

After the announcement, the rollout ran into trouble. On August 13, the food site Inspired Taste reported finding AI-generated step-by-step images inside an AI Overview that summarized one of its own recipes.27 Google responded four days later. It called the feature a small experiment that it had stopped running, and it did not say how widely the images had appeared or in which markets.27

The tracker's assessment is the most useful reading available: that rollout was stopped, but Google has not dropped the plan to bring image generation into AI Overviews.27 The underlying problem is real. AI Overviews lists the publishers it draws on. A generated image placed in that box, with nothing marking it as Google's, can look like the cited site's own work.27 For recipe and how-to publishers, whose business depends on original photos, that is more than a cosmetic issue.

The timing adds to the risk. The EU AI Act's transparency rules took effect on August 2, 2026. They require providers of systems that generate images, text, audio or video to mark that output in a machine-readable format.27 Google says every Nano Banana image carries an invisible SynthID watermark.26 Nano Banana 2.1 also supports C2PA Content Credentials on Google Cloud.6 Machine-readable marks don't fix the problem of what readers actually see, though. A watermark a reader can't see doesn't tell them who made the picture.

The publisher episode reads to us as the real story behind the announcement. The question isn't whether Google can generate images in search results. It plainly can. The open question is whether it can do that without blurring who made what on the page.

The model race underneath

The search features sit on top of a model family Google is updating quickly. Nano Banana launched in August 2025, Nano Banana Pro followed in November, and Nano Banana 2 arrived in February 2026. Nano Banana 2 became the default for image generation in the Gemini app, Google Lens and AI Mode across 141 countries.8 At I/O in May 2026, Sundar Pichai said the family had produced more than 50 billion images.4

Nano Banana 2.1 is built on Gemini 3.6 Flash.59 Google says it beats earlier versions "across the board," especially on visual design, mask-based editing and keeping subjects consistent.1 It handles up to 14 reference images, keeps track of four characters and ten objects, and outputs up to 4K in aspect ratios as wide as 8:1.57 The price has also fallen. Developer API image costs were cut in half, to $0.0336 for a 1K image and $0.0756 for 4K.5 Google released 2.1 across the Gemini app, AI Mode, AI Studio, Flow, Stitch, Google Ads and its enterprise platform at the same time.6 Nano Banana 2 shuts down in the API on October 29.6 The older Imagen models have already been retired in favor of Nano Banana.23

Outside opinions of 2.1 are more mixed than Google's. Google's own preference tests put it ahead of both Nano Banana 2 and Nano Banana Pro.9 On the public Arena leaderboard, though, it ranked fifth for text-to-image and sixth for editing as of launch day, behind OpenAI's GPT Image models.6 The Decoder found that Nano Banana Pro still produced more natural-looking images side by side, and that 2.1 sometimes got scale wrong.9 Google's message to search users is that it has the best model. Independent testing points to a cheaper, faster model that is competitive but not clearly the leader.

Why video belongs in this story

The search update is about still images, but the same strategy is playing out in video. In the Gemini app, Google is replacing Veo with Gemini Omni, a model that both generates and edits video.17 Gemini Omni Flash, introduced at I/O 2026, combines text, images, audio and video in one output and lets users make changes conversationally. It is limited to roughly 10-second clips.12 On the Artificial Analysis leaderboard for text-to-video with audio, Omni Flash ranks first, while Veo 3.1 sits in 14th place.16

Veo hasn't disappeared. Veo 3.1 is still the mature option for enterprise customers who want a longer track record and Google Cloud's compliance tools.16 Google is still pushing developers to move on, however. The Veo 3.1 preview endpoints in the Gemini API shut down on October 22, and Google recommends an Omni Flash model as the replacement.19 Google has not announced a Veo 4.19

The pattern is the same across both media. Google is folding separate, single-purpose models into multimodal Gemini-branded systems and then building them into products people already use. Imagen gave way to Nano Banana. Veo is giving way to Omni. Nano Banana 2.1 shipped across seven products at once. One analysis argued that putting the model into Ads, Search and enterprise tools all together makes it infrastructure, not a standalone app competing for downloads.4 Gemini itself already offers narrated 30-to-60-second Video Overviews and zoomable image explanations for some topics.22 Search, too, is starting to answer with pictures and visuals instead of only text.27

The takeaway

Google's image-generation push in AI Overviews looks like a sound strategic bet that got ahead of the work needed to make it acceptable to publishers. The redesigned Images gallery is a fairly safe change. Generating images inside search results is not. When a search engine creates content instead of pointing to it, the line between what the web published and what Google made becomes easier to blur. Google paused the August test but kept the plan. With image models updating roughly every few weeks and video heading the same way, how these features are labeled and credited in search results will matter as much as how good the models are.