於 Cycles 渲染器整合鏡面流形採樣以增強鏡面光傳輸

Abstract

寫實渲染必須精確模擬複雜的光傳輸現象，尤其是由鏡面反射與折射所產生的多次反彈焦散。標準路徑追蹤渲染器（如 Blender Cycles）在有效捕捉此類效果時往往效率不足。Cycles 雖於近期整合「流形下一事件估計」（Manifold Next Event Estimation, MNEE），得以處理特定折射型焦散，但該方法僅適用於投射陰影內且假設表面光滑，對陰影外的折射或反射焦散以及複數鏡面路徑場景仍然無能為力。本論文透過將「鏡面流形採樣」（Specular Manifold Sampling, SMS）整合至 Cycles，增強其折射焦散渲染能力，並記錄在處理反射交互作用時所面臨的主要挑戰。

研究首先擴充 Cycles 的流形計算流程，使之同時支援鏡面反射與折射情境，並實作無偏與有偏兩種 SMS 估計器。隨機初始化階段在鏡面曲面上取樣隨機點與微表面法向，以啟動流形探索並尋得多重光路；對應的機率估測機制分別透過重複試算（無偏）與固定解集合（有偏）完成，並可省略部分直接光照取樣流程。

本實作於自製基準場景中評估，使用渲染時間、每像素取樣數（spp）、均方根誤差（RMSE）與結構相似度指標（SSIM）等度量，與標準 Cycles 路徑追蹤及原始 MNEE 版本比較。結果顯示，整合 SMS 的 Cycles 能有效重現多次反彈折射焦散，並在相同比樣率下顯著降低噪點與誤差。

本研究最終開發出一套可處理多次反彈折射焦散的 Cycles 渲染器，並完整紀錄將先進研究移植至生產級渲染器時理論與實務落差的經驗，為後續鏡面光傳輸研究與開發奠定基礎。

Photorealistic rendering demands accurate simulation of complex light-transport phenomena, such as multi-bounce caustics generated by specular reflection and refraction. Standard path-tracing renderers, including Blender's Cycles, often struggle to efficiently capture these effects. While Cycles recently integrated Manifold Next Event Estimation (MNEE) to resolve specific refractive caustics, the technique remains limited to caustics that fall inside the caster's shadow, assumes smooth-shaded surfaces, and fails to capture refractive or reflective caustics outside shadow regions, or handle scenes with multiple concurrent specular paths.

This thesis addresses these shortcomings by detailing the process of extending Cycles' framework with Specular Manifold Sampling (SMS), aiming to enhance its caustic rendering capabilities for refractive paths while documenting the significant challenges encountered, particularly with reflective interactions.

The objective was achieved by enhancing Cycles' manifold-computation routines to support both reflective and refractive scenarios. New modules implement biased and unbiased SMS estimators. A stochastic initialization phase selects random surface points and microfacet normals to seed manifold exploration, enabling discovery of multiple light paths. Corresponding probability-estimation mechanisms were implemented for both the unbiased estimator (via repeated trials) and the biased estimator (via a fixed budget of unique solutions). These modules allow to skip parts of the existing direct-illumination sampling strategy.The implementation was evaluated on custom benchmark scenes featuring challenging refractive caustics. The results were assessed using metrics such as rendering time, samples per pixel (spp), root-mean-square error (RMSE), and Structural Similarity Index Measure (SSIM) against baseline Cycles with standard path tracing, and its original MNEE implementation.

The primary outcome is a Cycles renderer with new capabilities for multi-bounce refractive caustics via SMS, alongside a documented account of the practical difficulties in porting such advanced research, highlighting areas where theoretical promise meets implementation complexities.