Adaptive Guidance for Real-Time Pathtracer Denoising Precision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current real-time denoisers and filtering parametrizations, such as spatio-temporal variance guided filters, tend to over-blur images in scenarios where precision is required, leading to suboptimal results in computer-aided designs (CADs), while machine learning-based techniques are computationally expensive.
Innovation Solution
An adaptive guidance algorithm is implemented to control variance-based filtering, using a guidance map that considers features like surface normals, roughness, and material type to prevent over-blurring, especially in complex illumination scenarios, and is specialized for materials like sub surface scattering and hair.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If spatio-temporal variance guided filtering is used for real-time denoising, then convergence speed is improved, but image precision deteriorates due to over-blurring
Solution Approach 1:
The patent applies local quality by creating a guidance map that assigns different filtering strengths to different regions of the image. The guidance map is generated using features such as surface normals, roughness, and material type, allowing the filter to be aggressive in noisy regions while preserving details in important regions like edges and material boundaries. This resolves the contradiction by making the filtering adaptive rather than uniform.
Solution Approach 2:
The patent implements dynamics by making the filtering strength dynamic and adaptive based on local image characteristics. The guidance map dynamically adjusts the filtering parameters for each pixel based on features like variance, surface normals, and material properties. This allows the system to maintain high convergence speed while preserving image precision through adaptive, rather than static, filtering.
2Object-generated harmful factors
If aggressive variance-based filtering is applied to reduce noise, then denoising effectiveness is improved, but detail preservation worsens due to over-blurring of edges and features
Solution Approach 1:
The guidance map enables local quality control by differentiating between regions that need aggressive filtering (noisy uniform areas) and regions that need gentle filtering (edges, material boundaries, high-frequency areas). Features like surface normals and roughness help identify these regions, allowing the filter to reduce noise effectively while preserving important details.
Solution Approach 2:
The guidance map acts as an intermediary between the raw image and the filtering operation. It mediates the filtering strength by providing per-pixel or per-region guidance values that modulate the variance-based filtering. This intermediary layer allows the system to achieve both noise reduction and detail preservation by controlling the filtering behavior through the guidance map.
3Productivity
If real-time rendering is performed at high noise levels to accelerate processing, then rendering speed is improved, but image quality deteriorates requiring stronger filtering
Solution Approach 1:
The patent applies preliminary action by pre-computing the guidance map using features like surface normals, roughness, and material type before the actual denoising operation. This preliminary guidance information allows the subsequent filtering to be highly effective at lower strengths, enabling the system to render at high noise levels and then efficiently clean up the image without excessive blurring.
Solution Approach 2:
The patent uses parameter changes by adjusting the filtering parameters based on the guidance map and local image characteristics. The guidance map provides parameter modulation that adapts the filtering strength to the local content, allowing aggressive filtering where needed and gentle filtering where details matter. This enables high-quality results from high-noise renders.
Data Source
AI summary
The present embodiments generally relate to enhancing a quality of denoising in a real-time pathtracer. A pathtracer can implement a Monte-Carlo process, which can require many samples to converge to an error-acceptable result. Denoisers are often used in real-time (or near real-time), in order to accelerate the convergence process. The present embodiments further comprise systems and methods relating to an adaptive guidance algorithm to be implemented on top of various denoising algorithms (e.g., SVGF/ASVGF). The systems and methods as described herein can improve the gradient estimation quality, producing less blurry images. The systems and methods can also substantially assist many denoising algorithms in complicated illumination scenarios, which are common in CADs, such as indirect illuminated normal maps, specular paths, sharp specular reflections/refractions, very noisy direct illumination like dome lights, specular-specular sub-paths, sub surface scattering or hair, etc.