Spectral transforms for caustic denoising : a comparative analysis for Monte Carlo rendering

Abstract

This paper investigates the effectiveness of spectral transforms in denoising caustics in physically based Monte Carlo rendering, where preserving high-frequency details such as contours within caustics is crucial. We approach the challenge posed by noise and the complexity of caustics in rendering by integrating various search techniques, image similarity metrics, thresholding functions and spectral transforms to optimise thresholding coefficients. Our comparative analysis shows that spectral methods can outperform denoisers like ReBLUR and ReLAX in certain scenarios by preserving details within caustic patterns more effectively. However, AI-based denoisers generally deliver better overall noise reduction, albeit at the cost of losing some image details. While spectral transforms currently require offline processing and reference images to fine-tune denoising coefficients, initial results show promising potential for applying these coefficients to different scenes with similar characteristics, enhancing applicability.