Multi-Scale Generation of Spatial Interaction Scenes via Implicit Neural Representations and Diffusion Models

Abstract

With the growth of virtual reality and smart cities for dynamic space generation, traditional methods have challenges in multi-scale modeling of complex interactive scenarios, and it is difficult to coordinate the consistency of global structure and local details with generative adversarial networks and variational autoencoders. Although implicit neural representation and diffusion models have potential in continuous spatial modeling and high-quality generation, their fusion and dynamic scene applications have not been fully explored. In this paper, a multi-scale generation framework is proposed: an implicit neural encoder with coordinate-semantic decoupling at the core (spatial coordinates are encoded by Fourier and output by a three-layer fully connected SIREN network) and a multi-resolution conditional diffusion model (50-100 steps of global coarse sampling, 200-300 steps of local fine sampling, and 4-level implicit and diffusion features are fused through a gating mechanism) with a dynamic gradient propagation mechanism (spatiotemporal joint loss multi-resolution pyramid LSTM). Timing module Physical constraint correction) to achieve macro and micro collaborative generation. Based on 1200 sets of urban scene and indoor scene datasets (4-level scale, multi-format and multi-annotation), the generation quality (FID decreased by 18.7%), multi-scale consistency (SSIM improved by 23.4%), and physical rationality (collision pass rate increased by 31.2%) were better than BIM GIS, NeRF and other baselines after training on Intel Xeon Gold CPU and NVIDIA A100 GPU (PyTorch 2.0). With the introduction of progressive sampling, single scene generation at 2560×1440 resolution takes only 4.3 seconds, which is 2.6 times faster than traditional diffusion. The ablation experiment verifies the key role of implicit coding and diffusion denoising coupling (LPIPS is 15.9%), and the physical rule compliance rate in the dynamic test is 92.7%, laying the foundation for the real-time construction of virtual and real fusion scenarios and smart city applications.