Personality inference from social media enables personalized applications such as recommender systems, digital mental health, and user-centered design. This study examines Instagram and the predictive power of two modalities: captions and images. We extracted linguistic features (hashtags, emojis, sentiment markers) and visual attributes (color distributions, object and face counts, CNN-based semantic embeddings). Using the Big Five NEO PI-R model as ground truth, experiments were conducted on a self-collected dataset of 323 Moroccan Instagram users (178 women, 145 men). Several regression algorithms were applied under three conditions: text-only, image-only, and multimodal fusion. Importantly, captions outperform images for most traits except Neuroticism, while multimodal fusion surpasses both approaches for Openness and Conscientiousness. These findings highlight the complementary nature of captions and images.