In 2001, Google launched Image Search. It was triggered by a very concrete need: users searching for "Jennifer Lopez green Versace dress" couldn't find satisfactory results in text search — the dress was so famous the whole internet was talking about it, but no text result let users "see" what it actually looked like.

Twenty-five years later, Google Images processes over 12 billion visual queries per month, and Google Lens recognizes over 10 billion images monthly. From "searching for pictures" to "searching with pictures" to "generating pictures," visual search has traversed a quarter-century of evolution.

Three Eras: Index → Understand → Generate

Google divides visual search's 25 years into three phases — a useful framework.

Phase 1 (2001-2015): Indexing era. The core problem here was "finding relevant pictures." Technically it relied mostly on text tags and page context — search "cat" and you'd find images where alt text or surrounding text contained "cat." Visual search in this era was essentially an extension of text search; images were just indexed objects, and computers didn't truly "understand" what was in them.

Phase 2 (2015-2022): Understanding era. Deep learning exploded around 2015, computer vision accuracy leaped forward every year. Google Lens launched in 2017, letting users point their camera at anything and search directly — photograph a flower to identify it, snap a restaurant to see reviews and menus, capture a math problem to get step-by-step solutions. The core of this phase was "understanding image content": computers didn't just know an image was tagged "cat" — they knew it was an orange tabby, sitting on a couch, next to a coffee mug, sunny outside the window.

Phase 3 (2022-present): Generation era. Multimodal foundation models transformed visual search from "finding images" to "creating them." Now when you type "a futuristic cat cafe" into Google Images, it doesn't just find related pictures — it directly generates brand-new images matching your description. Take a photo of your living room and say "change the sofa to mid-century modern," and it renders the renovated result directly. The boundary between search and creation has disappeared.

The Next 25 Years: Vision Becomes the Primary Interface

Looking back 25 years, what's most striking isn't the technical progress itself — it's that visual search's evolution maps directly onto the broader trajectory of human-computer interaction.

In the PC era, we interacted with computers by typing text on keyboards. In the mobile internet era, we touched screens with our fingers and took photos with cameras. In the AI era, vision is becoming the most natural interaction interface — we point at something and ask AI what it is, show AI a picture and say "I want it like this," have AI directly generate the images we imagine.

What will visual search look like in another 25 years? Google's direction is "immersive, conversational, creative" — you'll be able to seamlessly search, understand, modify, and generate visual content within a single scene, no app-switching required. For example, you're renovating: take a photo of the bare apartment, say "I want Japanese minimalist style," and AI directly renders the furnished result, every piece of furniture clickable to purchase. You're planning a trip: point at a map and say "I want somewhere quiet, by the ocean, good for surfing," and AI generates a photorealistic preview and itinerary directly.

Twenty-five years ago, people needed an image to understand "how beautiful Jennifer Lopez's dress actually was." Twenty-five years from now, AI can directly generate any image you can imagine. Vision has gone from "indexed content" to "interaction interface" to "creative medium" — and that transformation is just getting started.

See you tomorrow.