DietDelta: A Vision-Language Approach for Dietary Assessment via Before-and-After Images

A team of researchers including Gautham Vinod, Siddeshwar Raghavan, Bruce Coburn, and Fengqing Zhu has introduced DietDelta, a novel vision-language framework that revolutionizes dietary assessment through before-and-after food photography. Unlike traditional methods that rely on single pre-consumption images and provide only meal-level estimates, DietDelta analyzes paired images to determine what was actually consumed. The system uses natural language prompts to localize specific food items and estimate their weight directly from standard RGB images, eliminating the need for restrictive inputs like depth sensing, multi-view imagery, or explicit segmentation masks.

DietDelta employs a two-stage training strategy that first estimates food weight from individual images, then predicts consumption by calculating weight differences between the before-and-after pairs. This approach enables precise, food-item-level nutritional analysis rather than coarse meal-level estimates. The researchers evaluated their method on three publicly available datasets and demonstrated consistent improvements over existing approaches, establishing a strong baseline for before-and-after dietary image analysis. The framework's simplicity and reliance on standard smartphone photography make it particularly promising for real-world applications in precision nutrition and health monitoring.