UW-Madison's camroll Agent Answers Personal Photo Questions Using 234x Fewer Tokens

Your camera roll is probably the richest personal archive you own. Thousands of photos, spanning years of meals, trips, and faces -- yet asking a simple question like "What did I eat the day after I visited NASA?" is essentially impossible with today's tools. A new paper from researchers at the University of Wisconsin-Madison and Adobe Research tackles this problem head-on, introducing both a benchmark dataset and a purpose-built AI agent for conversational question answering over personal photo libraries.

The problem with your photo app

Surveys suggest the average person accumulates roughly 3,139 photos on their phone, and while 65% of users say they took those photos to reflect on later, more than half feel overwhelmed when trying to find a specific moment. Current tools like Google Photos or Apple Photos are largely organized chronologically and support only basic similarity search -- find photos by face, by place, or by keyword. These systems still largely operate as a retrieval module at surface level, such as face/object detection or keyword-based search.

The technical challenge runs deeper than UI design. A single HD photo costs 1-3k tokens, so a full camera roll of thousands of images can easily reach 1-10 million tokens. That blows past the context window of most models, and even when it fits, long-context understanding degrades as input grows. The alternative -- RAG (retrieval-augmented generation), which builds a searchable text database and pulls relevant snippets at query time -- has its own flaw: images are often treated as independent units, without incorporating personal context such as events and relationships, which leads to noisy retrieval.

Google has introduced Gemini with Google Photos, enabling responses grounded in personal photo collections, and Meta's Muse Spark supports connecting to personal albums or Facebook/Instagram posts. However, little academic work studies how MLLMs reason over long-horizon personal visual streams, where information is fragmented across time and context. This paper is the first serious attempt to fill that gap with a proper benchmark and a dedicated agent.

The camroll dataset

The final dataset, camroll, contains 50 users, 31,476 images, and 2,500 QA pairs drawn from two sources: an in-house collection of 30 real smartphone users, and a curated subset of the public YFCC-100M dataset. Each camera roll spans 2-6 years of real photos, and every image is timestamped to the second. The dataset covers photos taken across 25+ countries on 5 continents.