New AI Can Search Inside Images Using Just Your Words

Imagine being able to find any screenshot or photo on your computer just by typing a description of what's in it—even if you never tagged it. That's what a new AI model called EmbeddingGemma2 can do. It's a small program (about 330MB, smaller than many phone apps) that turns both images and text into numerical codes, then compares them to find matches. A user tested it on thousands of screenshots from visual novels and found it works surprisingly well for its size, especially when searching in different languages.

The way it works is like giving each image a unique fingerprint based on its content. When you type a search query, the AI creates a similar fingerprint for your words and looks for images with the closest match. In the test, the model could find images containing specific objects or actions, even when the query was in Korean and the image had no text. For example, it successfully found images with a 'triangular cloth hat' when given a Korean sentence describing it. The model took about 8.5 minutes to process 3,003 images, which is reasonable for a one-time setup.

However, there are limitations. The AI sometimes gets confused by visual traits like hair color or facial expressions—it might ignore them if the text description is strong. It also requires careful wording; for instance, 'blond' vs. 'blonde' can affect results. The tester notes that these quirks might be due to the model's small size, which trades some accuracy for speed and efficiency. Also, the test was not scientific—only a few sets of images were used—so results may vary in real-world scenarios.

Despite these caveats, this technology could be a game-changer for anyone who deals with large collections of images, like photographers, researchers, or just people with messy photo libraries. Instead of manually tagging each image, you could simply describe what you're looking for and let the AI do the work. It's a step toward more intuitive search that understands context and meaning, not just keywords.