Abstract: Monte Carlo-based ray tracing has been widely adopted in photorealistic rendering due to its physical accuracy, flexibility, and universa
lity. However, existing hardware limitations restrict real-time implementations to low sample-per-pixel rates, where Monte Carlo integration exhibits significant noise under sparse sampling conditions. This paper employs neural network techniques to address the trade-off between quality and performance, effectively resolving noise issues in Monte Carlo ray tracing with low sample counts. The proposed framework incorporates three core innovations: a multi-scale dilated convolution module that  enhances denoising quality by fusing local feature details with global contextual information with minimal parameter overhead; a temporal cross-attention mechanism that improves temporal coherence by adaptively reusing denoised historical frames; and a composite loss function combining reconstruction and temporal consistency terms to strengthen detail preservation and stability. Evaluated on the BMFR benchmark dataset, our method achieved an average 3.7% improvement in peak signal-to-noise ratio (PSNR) and a 23.5% reduction in root mean squared error (RMSE) compared to state-of-the-art approaches while maintaining comparable structure similarity index measure (SSIM) performance. The inference time meets real-time requirements.

Key words: Monte Carlo ray tracing, neural network, real-time ray tracing denoising, kernel prediction network