New AI Spots Tiny Details in Crowded Images Without Training

Have you ever tried to find one specific spec, like screen size or battery life, on a dense product page covered in photos and fine print? It’s easy to miss things, and standard AI tools often struggle too. They look at a whole photo or section and compress it into one big summary, which can blend different details together. The researchers call this mix-up “semantic dilution” — think of it like stirring all the ingredients of a recipe into one unidentifiable soup.

SAGE fixes that by working more like a detective. First, it breaks the image into individual pieces, such as a product name, a price tag, or a specification. Then it arranges those pieces into a map showing how they relate. When you ask a question, SAGE searches this map piece by piece, comparing only the relevant details instead of using everything at once. The best part? SAGE needs no extra training. It works right away with existing AI models, which makes it faster and cheaper for companies to adopt.

To prove SAGE works, the team created a test set called DEAR, using 1,055 real product-page questions that require comparing multiple details, like “Which of these two laptops has more storage?” SAGE correctly answered 84.9% of those tricky questions, beating current search methods. That means fewer wrong answers when you ask a voice assistant about a product or try to compare options from a complex receipt or travel itinerary.

This is early research, not yet a consumer app. But the potential is big: anything from online shopping and medical documents to airplane manuals could become easier to scan. If SAGE finds its way into everyday tools, you may never again have to open ten tabs just to compare two products.