ReGraph: New Dataset for Recipe Graph Generation from Food Images

Researchers have introduced ReGraph, a large-scale recipe graph dataset designed to enhance recipe generation from food images by representing procedural cooking knowledge explicitly. The dataset encodes ingredients, cooking actions, and tools as entities, with attributes describing ingredient state changes and typed relations capturing manipulation targets, destinations, and procedural ordering. This structured representation aims to address limitations in current Large Multimodal Models (LMMs), which often produce plausible textual descriptions without encoding process-level knowledge. ReGraph provides a basis for evaluating whether model outputs truly understand the structured transformation process of cooking. The dataset is described in a paper titled 'ReGraph: Learning to Generate Recipe Graphs from Food Images,' available on arXiv (arXiv:2608.06917). The work is significant for the fields of computer vision and natural language processing, offering a new resource for research in procedural understanding and multimodal learning.