Carsten Dachsbacher
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Carsten Dachsbacher has a total of 76 co-authors. Here are the top co-authors by number of shared publications:
- Johannes Hanika
Karlsruhe Institute of Technology Germany
10 shared publications - Tobias Rapp
Karlsruhe Institute of Technology
7 shared publications - Vincent Schüßler
Karlsruhe Institute of Technology Germany
4 shared publications - Christoph Peters
Karlsruhe Institute of Technology
3 shared publications - Max Piochowiak
Karlsruhe Institute of Technology, Germany
3 shared publications - Reiner Dolp
Adobe Research France
3 shared publications
Latest Publications
- Open Access
- 13 November 2026
- Published by Wiley in Computer Graphics Forum
- p. e70645
Refit Always —Precomputed Ensembles of Bounding Volume Hierarchies for Parametric Animations
Bounding Volume Hierarchies (BVHs) are an integral part of modern real‐time rendering and simulation systems. For animated geometry, however, the runtime cost of rebuilding the BVH every frame to match the deformed geometry is often prohibitive. State‐of‐the‐art systems therefore amortize costs in several ways that negatively impact BVH quality: multi‐level BVHs confine updates to dynamic geometry, fast builders are chosen among algorithms striking different balances between BVH build time and BVH quality, and the BVH topology is updated at a lower temporal frequency with most frames only refitting bounding volumes instead of rebuilding the entire hierarchy. In this work, we present a parametric BVH, which is able to represent parametric animations without an expensive BVH rebuild within the scarce frame budget, and without the usual quality degradation associated with a BVH refit. For this, we precompute a small set of BVH topologies that jointly cover the animation's pose space with low error. At runtime, we obtain a BVH for the current pose by first selecting a BVH from the precomputed set, followed by a refit. We evaluate the practicality of our method on a large set of skeletal animations from the Mixamo dataset [Ado26]. Our approach matches results of high‐quality BVH builders at the significantly lower runtime cost of a BVH refit. Compared to the common methodology of rebuilding the BVH each frame using a fast builder (e.g. LBVH), this equates to 1.25 × lower SAH costs with 2.4 × lower runtimes.
- 13 November 2026
- Published by Wiley in Computer Graphics Forum
- p. e70645
Open AccessRefit Always —Precomputed Ensembles of Bounding Volume Hierarchies for Parametric Animations
Bounding Volume Hierarchies (BVHs) are an integral part of modern real‐time rendering and simulation systems. For animated geometry, however, the runtime cost of rebuilding the BVH every frame to match the deformed geometry is often prohibitive. State‐of‐the‐art systems therefore amortize costs in several ways that negatively impact BVH quality: multi‐level BVHs confine updates to dynamic geometry, fast builders are chosen among algorithms striking different balances between BVH build time and BVH quality, and the BVH topology is updated at a lower temporal frequency with most frames only refitting bounding volumes instead of rebuilding the entire hierarchy. In this work, we present a parametric BVH, which is able to represent parametric animations without an expensive BVH rebuild within the scarce frame budget, and without the usual quality degradation associated with a BVH refit. For this, we precompute a small set of BVH topologies that jointly cover the animation's pose space with low error. At runtime, we obtain a BVH for the current pose by first selecting a BVH from the precomputed set, followed by a refit. We evaluate the practicality of our method on a large set of skeletal animations from the Mixamo dataset [Ado26]. Our approach matches results of high‐quality BVH builders at the significantly lower runtime cost of a BVH refit. Compared to the common methodology of rebuilding the BVH each frame using a fast builder (e.g. LBVH), this equates to 1.25 × lower SAH costs with 2.4 × lower runtimes.
- Open Access
- 5 August 2026
- Published by Wiley in Computer Graphics Forum
- p. e70534
Optimized and Aligned Anisotropic Monte Carlo Sampling Patterns
Path tracing uses Monte Carlo integration to solve the rendering equation by evaluating the integrand at random sampling points. The convergence rate of the error can be significantly improved by using correlated instead of random sampling, especially on smooth integrands. However, on integrands with discontinuities due to, e.g., occlusion, the improvement is less pronounced. Prior work has shown that the variance of the estimator is equal to the product of the power spectrum of the integrand and the expected power spectrum of the sampling pattern. Discontinuous integrands have anisotropic power spectra that exhibit high energies along the directions of the discontinuities, which need to match the low‐energy directions of the sampling pattern to reduce variance. However, existing anisotropic sampling patterns have at most two low‐energy directions. Therefore, we propose an optimization‐based algorithm to synthesize two‐dimensional correlated sampling patterns with spectra that have more than two low‐energy directions, leading to improved convergence behavior. Further, we propose a practical and sample‐efficient algorithm that estimates the directions of discontinuities in the power spectra of two‐dimensional integrands. We show that our algorithm can reliably estimate these directions, allowing us to align the low‐energy directions of anisotropic correlated sampling patterns with the predicted directions. We demonstrate in an offline path tracer with light source sampling that our aligned sampling patterns improve the convergence rate on two‐dimensional integrands with multiple discontinuities compared to existing anisotropic sampling patterns and thus reduce the error more quickly.
- 5 August 2026
- Published by Wiley in Computer Graphics Forum
- p. e70534
Open AccessOptimized and Aligned Anisotropic Monte Carlo Sampling Patterns
Path tracing uses Monte Carlo integration to solve the rendering equation by evaluating the integrand at random sampling points. The convergence rate of the error can be significantly improved by using correlated instead of random sampling, especially on smooth integrands. However, on integrands with discontinuities due to, e.g., occlusion, the improvement is less pronounced. Prior work has shown that the variance of the estimator is equal to the product of the power spectrum of the integrand and the expected power spectrum of the sampling pattern. Discontinuous integrands have anisotropic power spectra that exhibit high energies along the directions of the discontinuities, which need to match the low‐energy directions of the sampling pattern to reduce variance. However, existing anisotropic sampling patterns have at most two low‐energy directions. Therefore, we propose an optimization‐based algorithm to synthesize two‐dimensional correlated sampling patterns with spectra that have more than two low‐energy directions, leading to improved convergence behavior. Further, we propose a practical and sample‐efficient algorithm that estimates the directions of discontinuities in the power spectra of two‐dimensional integrands. We show that our algorithm can reliably estimate these directions, allowing us to align the low‐energy directions of anisotropic correlated sampling patterns with the predicted directions. We demonstrate in an offline path tracer with light source sampling that our aligned sampling patterns improve the convergence rate on two‐dimensional integrands with multiple discontinuities compared to existing anisotropic sampling patterns and thus reduce the error more quickly.
- Open Access
- 29 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Computer Graphics and Interactive Techniques
- Vol. 9 (4) , 1-20
Optimizing Vulkan Dispatch Schedules for Real-Time U-Net Denoising 53
Image denoising is fundamental to Monte Carlo rendering. Recently, real-time path-traced applications have become viable, relying heavily on efficient denoising. Modern denoisers are often based on neural networks, most commonly variants of the U-Net architecture. While tooling for development and training of custom neural networks is well established in Python, existing deployment and interop strategies typically rely on external inference frameworks, incur host synchronization, data movement, or heavyweight dependencies, making them impractical for tightly integrated, real-time rendering pipelines. We present a light-weight and portable ONNX to Vulkan conversion framework that is designed for efficient, fully GPU-resident deployment of U-Net-based networks. We employ graph-based and profile-guided optimizations to determine optimal Vulkan dispatch schedules, including block sizes, data layouts, and kernel fusion decisions. Our method includes effective pruning of the vast search space, significantly reducing compile times. Using the Open Image Denoise (OIDN) network, we demonstrate runtime improvements over established inference frameworks such as TensorRT.
- 29 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Computer Graphics and Interactive Techniques
- Vol. 9 (4) , 1-20
Open AccessOptimizing Vulkan Dispatch Schedules for Real-Time U-Net Denoising 53
Image denoising is fundamental to Monte Carlo rendering. Recently, real-time path-traced applications have become viable, relying heavily on efficient denoising. Modern denoisers are often based on neural networks, most commonly variants of the U-Net architecture. While tooling for development and training of custom neural networks is well established in Python, existing deployment and interop strategies typically rely on external inference frameworks, incur host synchronization, data movement, or heavyweight dependencies, making them impractical for tightly integrated, real-time rendering pipelines. We present a light-weight and portable ONNX to Vulkan conversion framework that is designed for efficient, fully GPU-resident deployment of U-Net-based networks. We employ graph-based and profile-guided optimizations to determine optimal Vulkan dispatch schedules, including block sizes, data layouts, and kernel fusion decisions. Our method includes effective pruning of the vast search space, significantly reducing compile times. Using the Open Image Denoise (OIDN) network, we demonstrate runtime improvements over established inference frameworks such as TensorRT.
- Open Access
Minimizing Ray Tracing Memory Traffic through Quantized Structures and Ray Stream Tracing
Memory bandwidth constraints continue to be a significant limiting factor in ray tracing performance, particularly as scene complexity grows and computational capabilities outpace memory access speeds. This paper presents a memory-efficient ray tracing methodology that integrates compressed data structures with ray stream techniques to reduce memory traffic. The approach implements compressed BVH and triangle representations to minimize acceleration structure size in combination with ray stream tracing to reduce traversal stack memory traffic. The technique employs fixed-point arithmetic for intersection tests for prospective hardware with tailored integer operations. Despite using reduced precision, geometric holes are avoided by leveraging fixed-point arithmetic instead of encountering the floating-point rounding errors common in traditional approaches. Quantitative analysis demonstrates significant memory traffic reduction across various scene complexities and BVH configurations. The presented 8-wide BVH ray stream implementation reduces memory traffic to only 18% of traditional approaches by using 8-bit quantization for box and triangle coordinates and directly ray tracing these quantized structures. These reductions are especially beneficial for bandwidth-constrained hardware environments such as mobile devices. This integrated approach addresses both memory bandwidth limitations and numerical precision challenges inherent to modern ray tracing applications.
Open AccessMinimizing Ray Tracing Memory Traffic through Quantized Structures and Ray Stream Tracing
Memory bandwidth constraints continue to be a significant limiting factor in ray tracing performance, particularly as scene complexity grows and computational capabilities outpace memory access speeds. This paper presents a memory-efficient ray tracing methodology that integrates compressed data structures with ray stream techniques to reduce memory traffic. The approach implements compressed BVH and triangle representations to minimize acceleration structure size in combination with ray stream tracing to reduce traversal stack memory traffic. The technique employs fixed-point arithmetic for intersection tests for prospective hardware with tailored integer operations. Despite using reduced precision, geometric holes are avoided by leveraging fixed-point arithmetic instead of encountering the floating-point rounding errors common in traditional approaches. Quantitative analysis demonstrates significant memory traffic reduction across various scene complexities and BVH configurations. The presented 8-wide BVH ray stream implementation reduces memory traffic to only 18% of traditional approaches by using 8-bit quantization for box and triangle coordinates and directly ray tracing these quantized structures. These reductions are especially beneficial for bandwidth-constrained hardware environments such as mobile devices. This integrated approach addresses both memory bandwidth limitations and numerical precision challenges inherent to modern ray tracing applications.
- Open Access
- 27 May 2025
- Published by Wiley in Computer Graphics Forum
- Vol. 44 (3)
Random Access Segmentation Volume Compression for Interactive Volume Rendering
Segmentation volumes are voxel data sets often used in machine learning, connectomics, and natural sciences. Their large sizes make compression indispensable for storage and processing, including GPU video memory constrained real‐time visualization. Fast Compressed Segmentation Volumes (CSGV) [PD24] provide strong brick‐wise compression and random access at the brick level. Voxels within a brick, however, have to be decoded serially and thus rendering requires caching of visible full bricks, consuming extra memory. Without caching, accessing voxels can have a worst‐case decoding overhead of up to a full brick (typically over 32.000 voxels). We present CSGV‐R which provide true multi‐resolution random access on a per‐voxel level. We leverage Huffman‐shaped Wavelet Trees for random accesses to variable bit‐length encoding and their rank operation to query label palette offsets in bricks. Our real‐time segmentation volume visualization removes decoding artifacts from CSGV and renders CSGV‐R volumes without caching bricks at faster render times. CSGV‐R has slightly lower compression rates than CSGV, but outperforms Neuroglancer, the state‐of‐the‐art compression technique with true random access, with 2× to 4× smaller data sets at rates between 0.648% and 4.411% of the original volume sizes.
- 27 May 2025
- Published by Wiley in Computer Graphics Forum
- Vol. 44 (3)
Open AccessRandom Access Segmentation Volume Compression for Interactive Volume Rendering
Segmentation volumes are voxel data sets often used in machine learning, connectomics, and natural sciences. Their large sizes make compression indispensable for storage and processing, including GPU video memory constrained real‐time visualization. Fast Compressed Segmentation Volumes (CSGV) [PD24] provide strong brick‐wise compression and random access at the brick level. Voxels within a brick, however, have to be decoded serially and thus rendering requires caching of visible full bricks, consuming extra memory. Without caching, accessing voxels can have a worst‐case decoding overhead of up to a full brick (typically over 32.000 voxels). We present CSGV‐R which provide true multi‐resolution random access on a per‐voxel level. We leverage Huffman‐shaped Wavelet Trees for random accesses to variable bit‐length encoding and their rank operation to query label palette offsets in bricks. Our real‐time segmentation volume visualization removes decoding artifacts from CSGV and renders CSGV‐R volumes without caching bricks at faster render times. CSGV‐R has slightly lower compression rates than CSGV, but outperforms Neuroglancer, the state‐of‐the‐art compression technique with true random access, with 2× to 4× smaller data sets at rates between 0.648% and 4.411% of the original volume sizes.
- Open Access
- 22 May 2025
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Computer Graphics and Interactive Techniques
- Vol. 8 (1) , 1-18
Real-Time Markov Chain Path Guiding for Global Illumination and Single Scattering
We present a lightweight and unbiased path guiding algorithm tailored for real-time applications with highly dynamic content. The algorithm demonstrates effectiveness in guiding both direct and indirect illumination. Moreover, it can be extended to guide single scattering events in participating media. Building upon the screen-space approach by Dittebrandt et al. [ 2023 ], the incident light distribution is represented as a von Mises-Fisher mixture model, which is controlled by a Markov chain process. To extend the procedure to world space, our algorithm uses a unique Markov chain architecture, which resamples Markov chain states from an ensemble of hash grids. We combine multi-resolution adaptive grids with a static grid, ensuring rapid state exchange without compromising guiding quality. The algorithm imposes minimal prerequisites on scene representation and seamlessly integrates into existing path tracing frameworks. Through continuous multiple importance sampling, it remains independent of the equilibrium distribution of Markov chain and hash grid resampling. We perform an evaluation of the proposed methods across diverse scenarios. Additionally, we explore the algorithm’s viability in offline scenarios, showcasing its effectiveness in rendering volumetric caustics. We demonstrate the application of the proposed methods in a path tracing engine for the original Quake game. The demo project features path traced global illumination and single scattering effects at frame rates over 30 FPS on NVIDIA’s GeForce 20 series or AMD’s Radeon RX 6000 series without upscaling.
- 22 May 2025
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Computer Graphics and Interactive Techniques
- Vol. 8 (1) , 1-18
Open AccessReal-Time Markov Chain Path Guiding for Global Illumination and Single Scattering
We present a lightweight and unbiased path guiding algorithm tailored for real-time applications with highly dynamic content. The algorithm demonstrates effectiveness in guiding both direct and indirect illumination. Moreover, it can be extended to guide single scattering events in participating media. Building upon the screen-space approach by Dittebrandt et al. [ 2023 ], the incident light distribution is represented as a von Mises-Fisher mixture model, which is controlled by a Markov chain process. To extend the procedure to world space, our algorithm uses a unique Markov chain architecture, which resamples Markov chain states from an ensemble of hash grids. We combine multi-resolution adaptive grids with a static grid, ensuring rapid state exchange without compromising guiding quality. The algorithm imposes minimal prerequisites on scene representation and seamlessly integrates into existing path tracing frameworks. Through continuous multiple importance sampling, it remains independent of the equilibrium distribution of Markov chain and hash grid resampling. We perform an evaluation of the proposed methods across diverse scenarios. Additionally, we explore the algorithm’s viability in offline scenarios, showcasing its effectiveness in rendering volumetric caustics. We demonstrate the application of the proposed methods in a path tracing engine for the original Quake game. The demo project features path traced global illumination and single scattering effects at frame rates over 30 FPS on NVIDIA’s GeForce 20 series or AMD’s Radeon RX 6000 series without upscaling.
- Open Access
- 18 April 2025
- Published by Wiley in Computer Graphics Forum
- Vol. 44 (2)
Neural Two‐Level Monte Carlo Real‐Time Rendering
We introduce an efficient Two‐Level Monte Carlo (subset of Multi‐Level Monte Carlo, MLMC) estimator for real‐time rendering of scenes with global illumination. Using MLMC we split the shading integral into two parts: the radiance cache integral and the residual error integral that compensates for the bias of the first one. For the first part, we developed the Neural Incident Radiance Cache (NIRC) leveraging the power of tiny neural networks [MRNK21] as a building block, which is trained on the fly. The cache is designed to provide a fast and reasonable approximation of the incident radiance: an evaluation takes 2–25 × less compute time than a path tracing sample. This enables us to estimate the radiance cache integral with a high number of samples and by this achieve faster convergence. For the residual error integral, we compute the difference between the NIRC predictions and the unbiased path tracing simulation. Our method makes no assumptions about the geometry, materials, or lighting of a scene and has only few intuitive hyper‐parameters. We provide a comprehensive comparative analysis in different experimental scenarios. Since the algorithm is trained in an on‐line fashion, it demonstrates significant noise level reduction even for dynamic scenes and can easily be combined with other noise reduction techniques.
- 18 April 2025
- Published by Wiley in Computer Graphics Forum
- Vol. 44 (2)
Open AccessNeural Two‐Level Monte Carlo Real‐Time Rendering
We introduce an efficient Two‐Level Monte Carlo (subset of Multi‐Level Monte Carlo, MLMC) estimator for real‐time rendering of scenes with global illumination. Using MLMC we split the shading integral into two parts: the radiance cache integral and the residual error integral that compensates for the bias of the first one. For the first part, we developed the Neural Incident Radiance Cache (NIRC) leveraging the power of tiny neural networks [MRNK21] as a building block, which is trained on the fly. The cache is designed to provide a fast and reasonable approximation of the incident radiance: an evaluation takes 2–25 × less compute time than a path tracing sample. This enables us to estimate the radiance cache integral with a high number of samples and by this achieve faster convergence. For the residual error integral, we compute the difference between the NIRC predictions and the unbiased path tracing simulation. Our method makes no assumptions about the geometry, materials, or lighting of a scene and has only few intuitive hyper‐parameters. We provide a comprehensive comparative analysis in different experimental scenarios. Since the algorithm is trained in an on‐line fashion, it demonstrates significant noise level reduction even for dynamic scenes and can easily be combined with other noise reduction techniques.
- Open Access
- 18 April 2025
- Published by Wiley in Computer Graphics Forum
- Vol. 44 (2)
View‐Dependent Visibility Optimization for Monte Carlo Volume Visualization
Compared to classic ray marching‐based approaches, Monte Carlo ray tracing for volume visualization can provide faster frame times through progressive rendering, improved image quality, and allows for advanced illumination models more easily. Techniques such as the view‐dependent optimization of visibility and illumination of important regions, however, have been formulated for ray marching and rely on stepwise sampling along rays, and are thus incompatible with free‐flight distance sampling of state‐of‐the‐art Monte Carlo methods. In this paper we derive such a view‐dependent optimization for Monte Carlo ray tracing where the visibility to the camera, the illumination and opacity of important regions is optimized for both single and multiple scattering rendering. For this we define a post‐interpolative importance function, introduce an efficient data structure to sample, approximate and optimize the integrated extinction along rays, and devise an efficient Monte Carlo estimator for interactive visualization. Our method enables view‐dependent visibility optimization with moderate memory overhead and unbiased, progressive Monte Carlo volume visualization. We demonstrate our method for various volume data sets as well as for data‐dependent and spatially‐dependent importance functions.
- 18 April 2025
- Published by Wiley in Computer Graphics Forum
- Vol. 44 (2)
Open AccessView‐Dependent Visibility Optimization for Monte Carlo Volume Visualization
Compared to classic ray marching‐based approaches, Monte Carlo ray tracing for volume visualization can provide faster frame times through progressive rendering, improved image quality, and allows for advanced illumination models more easily. Techniques such as the view‐dependent optimization of visibility and illumination of important regions, however, have been formulated for ray marching and rely on stepwise sampling along rays, and are thus incompatible with free‐flight distance sampling of state‐of‐the‐art Monte Carlo methods. In this paper we derive such a view‐dependent optimization for Monte Carlo ray tracing where the visibility to the camera, the illumination and opacity of important regions is optimized for both single and multiple scattering rendering. For this we define a post‐interpolative importance function, introduce an efficient data structure to sample, approximate and optimize the integrated extinction along rays, and devise an efficient Monte Carlo estimator for interactive visualization. Our method enables view‐dependent visibility optimization with moderate memory overhead and unbiased, progressive Monte Carlo volume visualization. We demonstrate our method for various volume data sets as well as for data‐dependent and spatially‐dependent importance functions.
- Open Access
- 9 August 2024
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Computer Graphics and Interactive Techniques
- Vol. 7 (3) , 1-19
ReSTIR Subsurface Scattering for Real-Time Path Tracing
Subsurface scattering is an important visual cue and in real-time rendering it is often approximated using screen-space algorithms. Path tracing with the diffusion approximation can easily overcome the limitations of these algorithms, but increases image noise. We improve its efficiency by applying reservoir-based spatiotemporal importance resampling (ReSTIR) to subsurface light transport paths. For this, we adopt BSSRDF importance sampling for generating candidates. Further, spatiotemporal reuse requires shifting paths between domains. We observe that different image regions benefit most from either reconnecting through the translucent object (reconnection shift), or one vertex later (delayed reconnection shift). We first introduce a local subsurface scattering specific criterion for a hybrid shift that deterministically selects one of the two shifts for a path. Due to the locality, it cannot always choose the most efficient shift, e.g. near shadow boundaries. Therefore, we additionally propose a novel sequential shift to combine multiple shift mappings: We execute subsequent resampling passes, each one using a different shift, which does not require to deterministically choose a shift for a path. Instead, resampling can pick the most successful shift implicitly. Our method achieves realtime performance and significantly reduces noise and denoising artifacts in regions with visible subsurface scattering compared to standard path tracing with equal render time.
- 9 August 2024
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Computer Graphics and Interactive Techniques
- Vol. 7 (3) , 1-19
Open AccessReSTIR Subsurface Scattering for Real-Time Path Tracing
Subsurface scattering is an important visual cue and in real-time rendering it is often approximated using screen-space algorithms. Path tracing with the diffusion approximation can easily overcome the limitations of these algorithms, but increases image noise. We improve its efficiency by applying reservoir-based spatiotemporal importance resampling (ReSTIR) to subsurface light transport paths. For this, we adopt BSSRDF importance sampling for generating candidates. Further, spatiotemporal reuse requires shifting paths between domains. We observe that different image regions benefit most from either reconnecting through the translucent object (reconnection shift), or one vertex later (delayed reconnection shift). We first introduce a local subsurface scattering specific criterion for a hybrid shift that deterministically selects one of the two shifts for a path. Due to the locality, it cannot always choose the most efficient shift, e.g. near shadow boundaries. Therefore, we additionally propose a novel sequential shift to combine multiple shift mappings: We execute subsequent resampling passes, each one using a different shift, which does not require to deterministically choose a shift for a path. Instead, resampling can pick the most successful shift implicitly. Our method achieves realtime performance and significantly reduces noise and denoising artifacts in regions with visible subsurface scattering compared to standard path tracing with equal render time.
- Open Access
- 9 August 2024
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Computer Graphics and Interactive Techniques
- Vol. 7 (3) , 1-19
Optimizing Path Termination for Radiance Caching Through Explicit Variance Trading
Radiance caching allows to amortize the cost of path tracing by sharing contributions of path suffices in a spatial data structure. This sharing generally introduces bias, but it can be traded with variance by terminating into the cache at deeper path vertices. We develop a framework to implicitly reduce bias by optimizing path termination towards a chosen variance bound. Importantly, this bound can be chosen large enough while still being amenable to denoising. This results in longer paths being sampled with unbiased path tracing as permitted by the estimator variance and variance bound. To that end, we reformulate the variance of a path tracing estimator as a quantity that can be expressed locally for a path, relying on auxiliary statistics that are shared through the radiance cache structure independently of the path prefix. This allows to perform the optimization locally during path construction, while still translating to a global bound. Our method is capable of maintaining the variance bound in complex scenes, resulting in lower bias compared to other techniques such as heuristics based on ray differentials. We additionally present first findings on directly optimizing the bias-variance tradeoff based on local bias estimates of individual cache records, although the optimization is only approximate, resulting in suboptimal termination decisions.
- 9 August 2024
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Computer Graphics and Interactive Techniques
- Vol. 7 (3) , 1-19
Open AccessOptimizing Path Termination for Radiance Caching Through Explicit Variance Trading
Radiance caching allows to amortize the cost of path tracing by sharing contributions of path suffices in a spatial data structure. This sharing generally introduces bias, but it can be traded with variance by terminating into the cache at deeper path vertices. We develop a framework to implicitly reduce bias by optimizing path termination towards a chosen variance bound. Importantly, this bound can be chosen large enough while still being amenable to denoising. This results in longer paths being sampled with unbiased path tracing as permitted by the estimator variance and variance bound. To that end, we reformulate the variance of a path tracing estimator as a quantity that can be expressed locally for a path, relying on auxiliary statistics that are shared through the radiance cache structure independently of the path prefix. This allows to perform the optimization locally during path construction, while still translating to a global bound. Our method is capable of maintaining the variance bound in complex scenes, resulting in lower bias compared to other techniques such as heuristics based on ray differentials. We additionally present first findings on directly optimizing the bias-variance tradeoff based on local bias estimates of individual cache records, although the optimization is only approximate, resulting in suboptimal termination decisions.