DietAI24: comprehensive nutrition estimation using multimodal large language models
The food recognition engine inside NutriCamp is based on this paper. Instead of guessing nutrition from a model's memory, DietAI24 pairs a multimodal language model with Retrieval-Augmented Generation that grounds every estimate in the USDA Food and Nutrient Database for Dietary Studies (FNDDS).
On real-world mixed dishes the framework cuts food-weight and key-nutrient error sharply versus prior computer-vision methods, and reports far more than basic macros, so every number you see in the app traces back to an authoritative source you can audit.