STAR-NT: Spatiotemporal Acceleration of Real-Time Neural Transparency Rendering

2026-06-15

AI Summary

Neural order-independent transparency rendering incurs substantial computational overhead, making real-time performance challenging on resource-constrained devices. This work proposes a spatiotemporal acceleration framework that, for the first time, integrates spatially adaptive quadtree subdivision with temporal depth reprojection to significantly reduce rendering costs while preserving visual fidelity. By dynamically adjusting screen-space resolution and reusing information from previous frames, the method effectively exploits spatiotemporal coherence. The approach is designed for seamless integration into existing real-time rendering pipelines, offering an efficient solution for neural transparent rendering on mobile and legacy hardware.

Abstract

Neural order-independent transparency delivers high-quality rendering of overlapping transparent surfaces, but its geometry passes and network input generation remain costly, particularly on mobile and legacy hardware. We present a spatiotemporal acceleration framework that exploits spatial and temporal coherence to reduce this overhead while preserving visual quality. Spatially, we use adaptive quadtree-based screen-space subdivision to scale geometry pass resolution according to local color variance. Temporally, selected frames reuse the previous transparency result through depth-based reprojection instead of full rendering. Together, these optimizations reduce rendering cost and integrate efficiently into existing real-time rendering pipelines.

Problem

• neural transparency
• order-independent transparency
• real-time rendering
• rendering overhead
• mobile hardware

Innovation

• spatiotemporal acceleration
• neural transparency rendering
• adaptive quadtree subdivision
• depth-based reprojection
• real-time rendering