Hardware-Aware Quantized Generative Radiance Fields for Edge-Based Architectural Heritage Reconstruction

Abstract

Architectural heritage preservation increasingly relies on neural rendering techniques such as Neural Radiance Fields (NeRF) for high-fidelity 3D reconstruction. However, the computational demands of these models limit their deployment on resource-constrained edge devices commonly used in field documentation. This study investigates the feasibility of hardware-aware quantized generative radiance fields for edge-based architectural heritage reconstruction. A co-designed framework integrating quantization-aware training (8-, 6-, 4-, and 2-bit), multi-scale perceptual loss, and hardware-specific inference optimization was evaluated on six architecturally diverse heritage datasets. Reconstruction quality was assessed using PSNR, SSIM, LPIPS, edge preservation, texture analysis, and expert evaluation, while computational efficiency was measured through latency, memory usage, and energy consumption on representative edge platforms. Results show that 8-bit quantization preserves reconstruction quality comparable to full precision while reducing memory requirements by four times and significantly improving energy efficiency. The 6-bit and 4-bit models maintain most diagnostically important architectural details, with the 4-bit configuration representing the practical lower limit for heritage applications. Quantization-aware training consistently outperformed post-training quantization, and perceptual loss improved the preservation of fine architectural features under moderate compression. The proposed framework achieved near workstation-level reconstruction quality on dedicated edge accelerators while dramatically reducing computational cost, demonstrating the practicality of edge-based neural rendering for architectural heritage documentation and supporting broader adoption of efficient digital preservation technologies.