This paper introduces TuniFood-Bio, a bi-modal dataset combining high-resolution images of Tunisian dishes annotated via YOLO with structured nutritional metadata (calories, macronutrients) in tabular format. Our approach circumvents fine-grained ingredient extraction difficulties by linking recipe-level visual detection directly to a validated nutritional lookup table. Unlike joint bi-modal learning systems, only the image classifier is trained; nutritional values are retrieved without additional modeling. Benchmark results demonstrate robust performance (90.5% mAP@50) and exceptional classification precision (91.2%) under diverse real-world complexities—including variable lighting, ingredient occlusions, and regional dish variations—significantly outperforming unimodal baselines in challenging scenarios like overlapped stews (e.g., Chakchouka) and enclosed pastries (e.g., Brik). This late-stage integration improves classification accuracy while enabling practical nutrition-sensitive applications like calorie reporting and culturally adapted dietary tools. The dataset and pipeline are publicly available to advance food computing and regional health analysis.