用於蒙地卡羅渲染的雙歷史時空濾波方法：以變異數引導的時空濾波為例

Abstract

透過時空重用增加有效採樣數量對於實時光線追踪的降噪至關重要。現有的方法通常假設輸入是由固定的每像素樣本數生成的，並在時間重用當中使用指數移動平均來合併顏色。然而，當每像素樣本數變化時，這些方法無法有效利用由高樣本數生成的顏色。在這項研究中，我們提出了一種新方法，稱為「雙歷史方法」，以增強現有的時空濾波器，提升其處理可變樣本數輸入的能力。我們使用這種方法改進了變異數引導的時空濾波（SVGF）算法，並通過簡單的注視點渲染管線評估其性能。

Spatio-temporal reusing samples is critical to reducing noise for realtime ray-tracing. Existing methods usually assume inputs are generated by constant sample-per-pixel and use exponential moving averages to aggregate colors for temporal reuse. However, when sample-per-pixel vary, these approaches fail to effectively utilize colors generated by high sample counts. In this study, we proposed a novel method termed the "Two-history Approach" to augment existing spatio-temporal denoisers, enhancing their ability to handle inputs with variable sample count. We adapt the SVGF algorithm using our approach and evaluate its performance with a simple foveated rendering pipeline.