In the era of globalization and escalating sustainability demands, multi-modal supply chains, integrating road, rail, air, and sea transport offer significant efficiency potential but face challenges in coordination, freight exchange, and adaptive routing. This research introduces a generative AI framework based on a hybrid Variational Autoencoder-Generative Adversarial Network (VAE-GAN) architecture to enhance intelligent logistics optimization, enabling scenario generation for cost-effective, resilient, and eco-friendly operations. The framework is developed using Python, leveraging synthetic datasets generated to simulate realistic logistics scenarios, including emissions, disruptions, and multi-leg routes. A mixed-methods approach incorporated model training with PyTorch, discrete-event simulations for validation, and benchmarking against Mixed-Integer Linear Programming (MILP) and heuristic baselines. Evaluations focused on quantitative metrics such as total cost, transit time, CO \(_{2}\) emissions, and constraint satisfaction, alongside visual analyses of latent spaces and distributions. Results demonstrate that the VAE-GAN achieves near-optimal logistics costs, VAE-GAN, with superior computational efficiency—reducing execution times significantly compared to MILP—while maintaining scenario diversity and realism across baseline, disruption, high-demand, and sustainability conditions. The model exhibits zero constraint violations, stable prediction errors in feedback loops, and realistic trade-offs between cost and emissions. This research bridges a gap in applying generative AI to multi-modal logistics, improving flexibility, delivery efficiency, and ecological impact. Future extensions include real-time API integration and reinforcement learning for adaptive deployment.