Neural Rendering Meets Ray Tracing Pipelines: Integration of Denoising and Upscaling

While Monte Carlo ray tracing stands as the primary tool for photorealistic image synthesis, achieving noise-free results in real-time remains an essential challenge. Recent advances in hardware processing power, combined with artificial intelligence (AI) based algorithms, have significantly expanded the capabilities of this method. Originally restricted to offline rendering, ray tracing is now increasingly viable in interactive applications. Our work proposes a joint deep learning denoising and upscaling architecture operating in both irradiance and color spaces with geometry-guided attention, enabling real-time performance when combined with a Monte Carlo ray racer. As our primary contribution, we present a detailed analysis of how these techniques can be integrated into a ray tracing rendering pipeline, emphasizing key implementation challenges. By jointly exploring and synthesizing various design choices, optimizations, and features, we maximize both image quality and performance. Experimental results demonstrate that the proposed method effectively enhances image quality while maintaining real-time frame rates. We further show that such an architecture can be successfully trained on consumer-level hardware, lowering the barrier to developing competitive solutions. Ultimately, this work sheds light on the proprietary denoising and upscaling techniques employed by NVIDIA and AMD, thereby contributing to a deeper understanding of this rapidly evolving domain.