The Next Frontier of Generative Intelligence in Smarter Eco-Cities: A Systematic Review of Deep Generative and Hybrid Modeling Applications for Environmental Sustainability

• Provides the first cross-domain systematic synthesis of generative, hybrid, and mixed architectures for smarter eco-city intelligence. • Reveals distinct functional roles of GANs, VAEs, DMs, and transformers across urban–environmental systems. • Identifies a methodological shift from standalone models to integrated, multi-layered architectures enabling cross-system intelligence. • Proposes a novel hierarchical conceptual framework linking model capabilities to urban–ecological applications and decision-making layers. • Highlights technical and architectural constraints, data-related barriers, environmental and computational costs, and ethical and societal limitations Recent advancements in GenAI have gained increasing prominence, unveiling transformative opportunities to reconceptualize, analyze, manage, and plan urban–ecological systems. This development aligns with the emerging paradigm of smarter eco-cities, which function as experimental environments for testing and scaling technological innovations and environmental reforms. Despite the growing interest in applying GenAI to environmental sustainability and smart cities in recent years, the existing literature remains fragmented, domain-specific, and largely exploratory, often focusing on isolated use cases, individual model types, or narrowly defined functional areas. To the best of our knowledge, no review has systematically examined how Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), Diffusion Models (DMs), Transformers, and hybrid and mixed architectures can be adapted to the operational and strategic demands of smarter eco-cities. This study addresses this gap by providing the first systematic, cross-domain review of state-of-the-art applications of deep generative, hybrid, and mixed modeling for advancing environmental sustainability goals in smarter eco-cities. Methodologically, it synthesizes evidence from 145 studies through a unified framework that integrates aggregative, configurative, tabulated, and narrative approaches. Findings reveal that these modeling architectures advance generative, predictive, adaptive, and integrative intelligence across multiple urban–environmental domains, including energy systems, renewable energy, decarbonization, buildings, mobility and transport, waste and wastewater, air quality, carbon capture, water and hydrology, biodiversity and ecosystems, as well as disaster response and climate resilience. GANs excel in high-fidelity scenario realism and data augmentation; VAEs offer probabilistic interpretability, uncertainty modeling, and data imputation; DMs specialize in stable, controllable, and physically consistent simulations of complex spatiotemporal systems; Transformers enable multimodal temporal reasoning and long-range dependency modeling; and hybrid and mixed architectures combine these strengths or integrate them with discriminative, deterministic, heuristic, and/or symbolic components to support multifunctional, cross-system solutions. Building on these insights, the study proposes a novel conceptual framework that hierarchically integrates deep generative, hybrid, and mixed modeling approaches to advance smarter eco-city intelligence. However, persistent challenges remain, including technical and architectural constraints, hallucinations and epistemic uncertainty, data-related barriers, environmental and computational costs, and ethical and societal limitations, all of which hinder responsible and sustainable deployment. The findings carry significant implications for urban planners, researchers, and policymakers, highlighting the need for transdisciplinary collaboration, standardized evaluation frameworks across domains, and guidance for operationalizing urban-ecological intelligence strategies while ensuring ethical and inclusive implementation.