Abstract: A fully-connected neural network may be configured for execution by a processor as a fully-fused neural network by limiting slow global memory accesses to reading and writing inputs to and outputs from the fully-connected neural network. The computational cost of fully-connected neural networks scale quadratically with its width, whereas its memory traffic scales linearly. Modern graphics processing units typically have much greater computational throughput compared with memory bandwidth, so that for narrow, fully-connected neural networks, the linear memory traffic is the bottleneck. The key to improving performance of the fully-connected neural network is to minimize traffic to slow “global” memory (off-chip memory and high-level caches) and to fully utilize fast on-chip memory (low-level caches, “shared” memory, and registers), which is achieved by the fully-fused approach.

Abstract: Monte Carlo and quasi-Monte Carlo integration are simple numerical recipes for solving complicated integration problems, such as valuating financial derivatives or synthesizing photorealistic images by light transport simulation. A drawback of a straightforward application of (quasi-)Monte Carlo integration is the relatively slow convergence rate that manifests as high error of Monte Carlo estimators. Neural control variates may be used to reduce error in parametric (quasi-)Monte Carlo integration—providing more accurate solutions in less time. A neural network system has sufficient approximation power for estimating integrals and is efficient to evaluate. The efficiency results from the use of a first neural network that infers the integral of the control variate and using normalizing flows to model a shape of the control variate.

Abstract: Monte Carlo and quasi-Monte Carlo integration are simple numerical recipes for solving complicated integration problems, such as valuating financial derivatives or synthesizing photorealistic images by light transport simulation. A drawback of a straightforward application of (quasi-)Monte Carlo integration is the relatively slow convergence rate that manifests as high error of Monte Carlo estimators. Neural control variates may be used to reduce error in parametric (quasi-)Monte Carlo integration—providing more accurate solutions in less time. A neural network system has sufficient approximation power for estimating integrals and is efficient to evaluate. The efficiency results from the use of a first neural network that infers the integral of the control variate and using normalizing flows to model a shape of the control variate.

Abstract: A fully-connected neural network may be configured for execution by a processor as a fully-fused neural network by limiting slow global memory accesses to reading and writing inputs to and outputs from the fully-connected neural network. The computational cost of fully-connected neural networks scale quadratically with its width, whereas its memory traffic scales linearly. Modern graphics processing units typically have much greater computational throughput compared with memory bandwidth, so that for narrow, fully-connected neural networks, the linear memory traffic is the bottleneck. The key to improving performance of the fully-connected neural network is to minimize traffic to slow “global” memory (off-chip memory and high-level caches) and to fully utilize fast on-chip memory (low-level caches, “shared” memory, and registers), which is achieved by the fully-fused approach.

Abstract: A fully-connected neural network may be configured for execution by a processor as a fully-fused neural network by limiting slow global memory accesses to reading and writing inputs to and outputs from the fully-connected neural network. The computational cost of fully-connected neural networks scale quadratically with its width, whereas its memory traffic scales linearly. Modern graphics processing units typically have much greater computational throughput compared with memory bandwidth, so that for narrow, fully-connected neural networks, the linear memory traffic is the bottleneck. The key to improving performance of the fully-connected neural network is to minimize traffic to slow “global” memory (off-chip memory and high-level caches) and to fully utilize fast on-chip memory (low-level caches, “shared” memory, and registers), which is achieved by the fully-fused approach.

Abstract: A real-time neural radiance caching technique for path-traced global illumination is implemented using a neural network for caching scattered radiance components of global illumination. The neural (network) radiance cache handles fully dynamic scenes, and makes no assumptions about the camera, lighting, geometry, and materials. In contrast with conventional caching, the data-driven approach sidesteps many difficulties of caching algorithms, such as locating, interpolating, and updating cache points. The neural radiance cache is trained via online learning during rendering. Advantages of the neural radiance cache are noise reduction and real-time performance. Importantly, the runtime overhead and memory footprint of the neural radiance cache are stable and independent of scene complexity.

Abstract: A fully-connected neural network may be configured for execution by a processor as a fully-fused neural network by limiting slow global memory accesses to reading and writing inputs to and outputs from the fully-connected neural network. The computational cost of fully-connected neural networks scale quadratically with its width, whereas its memory traffic scales linearly. Modern graphics processing units typically have much greater computational throughput compared with memory bandwidth, so that for narrow, fully-connected neural networks, the linear memory traffic is the bottleneck. The key to improving performance of the fully-connected neural network is to minimize traffic to slow “global” memory (off-chip memory and high-level caches) and to fully utilize fast on-chip memory (low-level caches, “shared” memory, and registers), which is achieved by the fully-fused approach.

Abstract: A real-time neural radiance caching technique for path-traced global illumination is implemented using a neural network for caching scattered radiance components of global illumination. The neural (network) radiance cache handles fully dynamic scenes, and makes no assumptions about the camera, lighting, geometry, and materials. In contrast with conventional caching, the data-driven approach sidesteps many difficulties of caching algorithms, such as locating, interpolating, and updating cache points. The neural radiance cache is trained via online learning during rendering. Advantages of the neural radiance cache are noise reduction and real-time performance. Importantly, the runtime overhead and memory footprint of the neural radiance cache are stable and independent of scene complexity.

Abstract: Monte Carlo and quasi-Monte Carlo integration are simple numerical recipes for solving complicated integration problems, such as valuating financial derivatives or synthesizing photorealistic images by light transport simulation. A drawback of a straightforward application of (quasi-)Monte Carlo integration is the relatively slow convergence rate that manifests as high error of Monte Carlo estimators. Neural control variates may be used to reduce error in parametric (quasi-)Monte Carlo integration—providing more accurate solutions in less time. A neural network system has sufficient approximation power for estimating integrals and is efficient to evaluate. The efficiency results from the use of a first neural network that infers the integral of the control variate and using normalizing flows to model a shape of the control variate.

Abstract: Particular embodiments decompose an image comprising a scene into a diffuse component and a specular component. Each of the components represent a contribution to lighting in the scene. A set of motion vectors may be extracted in order to capture motion in the scene. Finally, a final contribution of each of the components to the image may be computed based on the motion vectors.

Abstract: Particular embodiments perform a light path analysis of an image comprising a scene, wherein the scene comprises at least one refractive or reflective object. The image may be decomposed based on the light path analysis into a plurality of components, each of the components representing a contribution to lighting in the scene by a different type of light interaction. For each of the components, one or more motion vectors are extracted for each of the components in order to capture motion in the scene. Finally, a final contribution of each of the components to the image is computed based on the motion vectors.

Abstract: According to one implementation, a system includes a computing platform having a hardware processor and a system memory storing a software code including multiple artificial neural networks (ANNs). The hardware processor executes the software code to partition a multi-dimensional input vector into a first vector data and a second vector data, and to transform the second vector data using a first piecewise-polynomial transformation parameterized by one of the ANNs, based on the first vector data, to produce a transformed second vector data. The hardware processor further executes the software code to transform the first vector data using a second piecewise-polynomial transformation parameterized by another of the ANNs, based on the transformed second vector data, to produce a transformed first vector data, and to determine a multi-dimensional output vector based on an output from the plurality of ANNs.

Abstract: According to one implementation, an image rendering system includes a computing platform having a hardware processor and a system memory storing an image denoising software code. The hardware processor executes the image denoising software code to receive an image file including multiple pixels, each pixel containing multiple depth bins, and to select a pixel including a noisy depth bin from among the pixels for denoising. The hardware processor further executes the image denoising software code to identify a plurality of reference depth bins from among the depth bins contained in one or more of the pixels, for use in denoising the noisy depth bin, and to denoise the noisy depth bin using an average of depth bin values corresponding respectively to each of the reference depth bins.

Abstract: According to one implementation, a system includes a computing platform having a hardware processor and a system memory storing a software code including multiple artificial neural networks (ANNs). The hardware processor executes the software code to partition a multi-dimensional input vector into a first vector data and a second vector data, and to transform the second vector data using a first piecewise-polynomial transformation parameterized by one of the ANNs, based on the first vector data, to produce a transformed second vector data. The hardware processor further executes the software code to transform the first vector data using a second piecewise-polynomial transformation parameterized by another of the ANNs, based on the transformed second vector data, to produce a transformed first vector data, and to determine a multi-dimensional output vector based on an output from the plurality of ANNs.

Abstract: A video rendering system includes a field-of-view detector, a display, and a computing platform including a hardware processor and a memory storing a multi-viewpoint video rendering software code. The hardware processor executes the multi-viewpoint video rendering software code to parameterize visible surfaces in a scene to define multiple texels for each visible surface, precompute one or more illumination value(s) for each texel of each visible surface, and for each texel of each visible surface, store the illumination value(s) in a cache assigned to the texel. In addition, the multi-viewpoint video rendering software code receives a perspective data from the field-of-view detector identifying one of multiple permissible perspectives for viewing the scene, and renders the scene on the display in real-time with respect to receiving the perspective data, based on the identified perspective and using one or more of the illumination value(s) precomputed for each texel of each visible surface.

Abstract: The disclosure provides an approach for denoising (also referred to as “filtering”) rendered images. In one embodiment, a denoising application takes as input rendered images and feature buffers that encode image information such as surface positions, surface depths, surface normals, surface albedos, and distances to the camera. For each pixel in a received image, the denoising application performs a first-order regression in a predefined neighborhood of the pixel to find a linear combination of pixel features that fits pixel colors in the predefined neighborhood. In such a first-order regression, the local regression weight of each pixel in the neighborhood may be determined using a metric which computes distances based on color values in patches around pixels being compared. In another embodiment, collaborative filtering may be performed in which filtered output from the first-order regression in each neighborhood is averaged with filtered output from overlapping neighborhoods to obtain a final output.

Abstract: According to one implementation, an image rendering system includes a computing platform having a hardware processor and a system memory storing an image denoising software code. The hardware processor executes the image denoising software code to receive an image file including multiple pixels, each pixel containing multiple depth bins, and to select a pixel including a noisy depth bin from among the pixels for denoising. The hardware processor further executes the image denoising software code to identify a plurality of reference depth bins from among the depth bins contained in one or more of the pixels, for use in denoising the noisy depth bin, and to denoise the noisy depth bin using an average of depth bin values corresponding respectively to each of the reference depth bins.

Abstract: The disclosure provides an approach for denoising (also referred to as “filtering”) rendered images. In one embodiment, a denoising application takes as input rendered images and feature buffers that encode image information such as surface positions, surface depths, surface normals, surface albedos, and distances to the camera. For each pixel in a received image, the denoising application performs a first-order regression in a predefined neighborhood of the pixel to find a linear combination of pixel features that fits pixel colors in the predefined neighborhood. In such a first-order regression, the local regression weight of each pixel in the neighborhood may be determined using a metric which computes distances based on color values in patches around pixels being compared. In another embodiment, collaborative filtering may be performed in which filtered output from the first-order regression in each neighborhood is averaged with filtered output from overlapping neighborhoods to obtain a final output.

Abstract: Particular embodiments decompose an image comprising a scene into a diffuse component and a specular component. Each of the components represent a contribution to lighting in the scene. A set of motion vectors may be extracted in order to capture motion in the scene. Finally, a final contribution of each of the components to the image may be computed based on the motion vectors.

Abstract: Particular embodiments perform a light path analysis of an image comprising a scene, wherein the scene comprises at least one refractive or reflective object. The image may be decomposed based on the light path analysis into a plurality of components, each of the components representing a contribution to lighting in the scene by a different type of light interaction. For each of the components, one or more motion vectors are extracted for each of the components in order to capture motion in the scene. Finally, a final contribution of each of the components to the image is computed based on the motion vectors.