Pic2Plate: A Vision-Language and Retrieval-Augmented Framework for Personalized Recipe Recommendations.

Choosing nutritious foods is essential for daily health, but finding recipes that match available ingredients and dietary preferences can be challenging. Traditional recommendation methods often lack personalization and accurate ingredient recognition. Personalized systems address this by integrating user preferences, dietary needs, and ingredient availability. This study presents Pic2Plate, a framework combining Vision-Language Models (VLMs) and Retrieval-Augmented Generation (RAG) to overcome these challenges. Pic2Plate uses advanced image recognition to extract ingredient lists from user images and RAG to retrieve and personalize recipe recommendations. Leveraging smartphone camera sensors ensures accessibility and portability. Pic2Plate's performance was evaluated in two areas: ingredient detection accuracy and recipe relevance. The ingredient detection module, powered by GPT-4o, achieved strong results with precision (0.83), recall (0.91), accuracy (0.77), and F1-score (0.86), demonstrating effectiveness in recognizing diverse food items. A survey of 120 participants assessed recipe relevance, with model rankings calculated using the Bradley-Terry method. Pic2Plate's VLM and RAG integration consistently outperformed other models. These results highlight Pic2Plate's ability to deliver context-aware, reliable, and diverse recipe suggestions. The study underscores its potential to transform recipe recommendation systems with a scalable, user-centric approach to personalized cooking.