Google Images is celebrating its 25th anniversary this week, and the team at Google dropped a retrospective that's part product launch, part victory lap. Two new features shipped—a personalized browseable homepage for Google Images and native image generation in AI Overviews—plus a timeline of 13 major milestones spanning two and a half decades.
What strikes me reading through this isn't the PR polish. It's how the core primitives of visual search have compounded into something that looks radically different from where it started. The journey from "show me Jennifer Lopez's green Versace dress" in 2001 to "use my live camera feed to have a video call with Search" in 2025 is genuinely wild.
Let's walk through what's new, what's interesting in the history, and what the visual fan-out technique tells us about where multimodal AI is headed.
The New Stuff: Browse and Generate
Google is rolling out two features over the coming weeks:
A dynamic, personalized Google Images homepage. Instead of a blank search box, you get an "immersive gallery of images from across the web—updated in real time and intelligently tailored to your unique interests." As you save ideas to collections, they appear as tabs above the gallery. Desktop-only, U.S.-only, English-only for now. You need to be signed in.
Image generation directly in AI Overviews. Using Google's latest Nano Banana model (yes, that's the real name they shipped in their Gemini model family post), you can now type a text prompt in Search and get a generated image when no existing web image fits. This launches in all regions that currently support image creation in AI Mode.
The personalized homepage is interesting positioning. Google is essentially admitting that search-as-intent-capture isn't always the right frame—sometimes you want to browse, Pinterest-style, and let the algorithm surface inspiration. The fact that they're doing real-time updates suggests they're running some kind of continuous ranker over your interaction history and the live web graph.
The image generation play is more straightforward: fill the gap when retrieval fails. The choice to embed this in AI Overviews rather than a separate tool is smart—it keeps the UX unified and signals that generation is just another Search result type, not a distinct product.
The Historical Arc: Three Eras of Visual Search
The retrospective covers 13 launches from 2001 to 2026. You can group them into three rough eras:
Era 1: Making the Web Visual (2001–2011)
- 2001: Google Images. Born because people wanted to see JLo's dress, not read about it. The original insight: text links aren't enough when the query is inherently visual.
- 2009: Similar Images. Query expansion by example. Click an image, get more like it. Simple, but it introduced the idea that images themselves could be queries.
- 2011: Search by Image. Upload an image or paste a URL, find its source or visually similar content. This inverted the whole flow—your starting point didn't have to be words.
Era 2: Multimodal Fusion (2018–2024)
- 2018: Google Lens in Search. Your phone camera becomes a search box. Point, tap, get results. No typing.
- 2022: Multisearch in Lens. Combine image + text in a single query. "What inspired this design?" with a photo of a landmark. This is where true multimodal fusion starts.
- 2024: Circle to Search. Gesture-based selection on Android screens. Circle, highlight, scribble, or tap anything on your display to search it without leaving your app. Now on over 580 million Android devices.
Era 3: AI-Native Visual Understanding (2025–2026)
- 2025: Lens + AI Mode. Multimodal Gemini processes entire scenes using "visual image fan-out"—breaking one image into dozens of sub-queries to understand full context.
- 2025: Search Live. Share your phone's live camera feed in a voice conversation with AI Mode. Video call with Search.
- 2025: Visual Results in AI Mode. Conversational visual exploration: "barrel jeans that aren't too baggy" → grid of shoppable products.
- 2026: Circle to Search Multi-Object Recognition. Fan-out again: analyze multiple objects in a single image simultaneously.
- 2026: Intelligent Search Box. Upload multiple images with detailed questions, get AI Mode responses.
The inflection point is clear: 2025 is when Google went all-in on visual fan-out and real-time multimodal reasoning. Everything before that is retrieval and matching. Everything after is generative understanding.
Visual Image Fan-Out: The Underrated Primitive
The "visual image fan-out" technique shows up three times in the timeline (Lens + AI Mode, Circle to Search multi-object, and implicitly in the intelligent search box). Google describes it as breaking "a single image search into dozens of sub-queries to understand the full visual context."
This is a genuinely clever architectural move, and it's worth unpacking why.
Traditional image search is an embedding problem: encode the image, encode the corpus, find nearest neighbors. You get one vector, one retrieval pass, one result set. It's fast and cheap, but it collapses all the information in a complex scene into a single point in latent space.
Fan-out inverts this. Instead of one dense query, you emit many targeted sub-queries:
- "What's the fabric of the shirt?"
- "What brand makes that shoe?"
- "What's the architectural style of the building in the background?"
- "Where was this photo taken?"
Each sub-query can be answered independently, then fused. This is much more expensive—dozens of retrieval or generation calls per image—but it's also much more expressive. You can capture multiple intents in a single user action.
The tradeoff is latency and cost. Google has the infrastructure to absorb this. Most startups don't. But as inference gets cheaper and models get faster, fan-out starts to look like the right default for any visual understanding task that isn't pure classification.
What This Tells Us About Multimodal AI
Three observations:
First, the camera is the new text box. Every major visual search feature since 2018 is camera-first. The intelligent search box even lets you upload multiple images at once. The shift from "describe what you want" to "show what you want" is basically complete.
Second, retrieval and generation are merging. The new image generation feature in AI Overviews isn't a separate product—it's a fallback when retrieval fails. Google is treating the boundary between "find it" and "make it" as a seamless continuum, not a product distinction.
Third, multimodal AI is embarrassingly parallel. Fan-out works because different parts of a scene are informationally independent. This is why visual understanding scaled faster than we expected—you can throw more compute at it and get linear returns.
Open Questions
A few things the post doesn't address:
How much does visual fan-out cost? If you're running dozens of sub-queries per image, what's the actual inference budget? And how much of this is cached or deduplicated across users?
What's the training data strategy for Nano Banana? They don't mention DALL·E 3 or Imagen, which suggests this might be a new architecture entirely. But Google has been opaque about synthetic data and RL fine-tuning for visual generation.
What happens to traditional image SEO? If Google is generating images and personalizing the homepage feed, does the open web still surface? Or are we heading toward a closed-loop model where Google answers visual queries internally?
The Bottom Line
Twenty-five years is a long time in internet years. The fact that Google Images started as a workaround for a celebrity dress and ended up as a real-time multimodal AI platform with live camera feeds and visual fan-out is a testament to compound iteration.
The new features are incremental, but the retrospective is a useful reminder of how far visual search has come. From text-only blue links to "have a video call with Search" is a hell of an arc.
And if you're building anything in the multimodal space, the fan-out technique is worth studying. It's not magic—it's just parallelism and fusion—but it's the kind of architectural choice that turns a demo into a product.