H-GIM: a human-centered generative immersive multimedia framework for adaptive immersive multimedia systems

Immersive multimedia systems based on Extended Reality (XR) are increasingly adopted in education, entertainment, and collaborative environments. However, traditional multimedia production pipelines rely heavily on manual content creation and static interaction models, limiting scalability and real-time adaptation to user behavior. This paper proposes H-GIM (Human-Centered Generative Immersive Multimedia), a framework powered by generative and multimodal artificial intelligence. The proposed architecture integrates multimodal foundation models capable of processing textual, audio, visual, and spatial signals through a router-based orchestration layer. To address the lack of large-scale XR training datasets, we introduce a synthetic data generation pipeline that automatically produces high-fidelity 3D assets using procedurally guided generative models. Importantly, 3-D asset generation is hybrid: geometry is produced via procedural seeds with physics validation, while SDXL-Turbo is used only for texture maps, ensuring physical plausibility without claiming end-to-end diffusion-based mesh generation. Furthermore, an affective interaction module dynamically adapts multimedia complexity based on user cognitive load and behavioral signals. Experimental evaluation on a functional prototype demonstrates that H-GIM achieves an average asset-generation time of 1.1 ± 0.2 min, a mean interaction latency of 42 ms, and a user-presence score of 0.81 ± 0.04. Results from a user study ( n = 50) further indicate statistically significant improvements in immersive experience compared with conventional XR workflows ( p < 0.001). These results demonstrate the potential of generative and multimodal AI to enable scalable, adaptive, and human-centered immersive multimedia systems.