We present a complete system for real-time rendering of scenes with complex appearance previously reserved for offline use. This is achieved with a combination of algorithmic and system level innovations. Our appearance model utilizes learned hierarchical textures that are interpreted using neural decoders, which produce reflectance values and importance-sampled directions. To best utilize the modeling capacity of the decoders, we equip the decoders with two graphics priors. The first prior—transformation of directions into learned shading frames—facilitates accurate reconstruction of mesoscale effects. The second prior—a microfacet sampling distribution—allows the neural decoder to perform importance sampling efficiently. The resulting appearance model supports anisotropic sampling and level-of-detail rendering, and allows baking deeply layered material graphs into a compact unified neural representation. By exposing hardware accelerated tensor operations to ray tracing shaders, we show that it is possible to inline and execute the neural decoders efficiently inside a real-time path tracer. We analyze scalability with increasing number of neural materials and propose to improve performance using code optimized for coherent and divergent execution. Our neural material shaders can be over an order of magnitude faster than non-neural layered materials. This opens up the door for using film-quality visuals in real-time applications such as games and live previews.

Abstract We introduce MesoGAN, a model for generative 3D neural textures. This new graphics primitive represents mesoscale appearance by combining the strengths of generative adversarial networks (StyleGAN) and volumetric neural field rendering. The primitive can be applied to surfaces as a neural reflectance shell; a thin volumetric layer above the surface with appearance parameters defined by a neural network. To construct the neural shell, we first generate a 2D feature texture using StyleGAN with carefully randomized Fourier features to support arbitrarily sized textures without repeating artefacts. We augment the 2D feature texture with a learned height feature, which aids the neural field renderer in producing volumetric parameters from the 2D texture. To facilitate filtering, and to enable end‐to‐end training within memory constraints of current hardware, we utilize a hierarchical texturing approach and train our model on multi‐scale synthetic datasets of 3D mesoscale structures. We propose one possible approach for conditioning MesoGAN on artistic parameters (e.g. fibre length, density of strands, lighting direction) and demonstrate and discuss integration into physically based renderers.

We present a method for reducing errors---variance and bias---in physically based differentiable rendering (PBDR). Typical applications of PBDR repeatedly render a scene as part of an optimization loop involving gradient descent. The actual change introduced by each gradient descent step is often relatively small, causing a significant degree of redundancy in this computation. We exploit this redundancy by formulating a gradient estimator that employs a recursive control variate , which leverages information from previous optimization steps. The control variate reduces variance in gradients, and, perhaps more importantly, alleviates issues that arise from differentiating loss functions with respect to noisy inputs, a common cause of drift to bad local minima or divergent optimizations. We experimentally evaluate our approach on a variety of path-traced scenes containing surfaces and volumes and observe that primal rendering efficiency improves by a factor of up to 10.

Abstract We investigate the use of neural fields for modelling diverse mesoscale structures, such as fur, fabric and grass. Instead of using classical graphics primitives to model the structure, we propose to employ a versatile volumetric primitive represented by a neural reflectance field (NeRF‐Tex), which jointly models the geometry of the material and its response to lighting. The NeRF‐Tex primitive can be instantiated over a base mesh to ‘texture’ it with the desired meso and microscale appearance. We condition the reflectance field on user‐defined parameters that control the appearance. A single NeRF texture thus captures an entire space of reflectance fields rather than one specific structure. This increases the gamut of appearances that can be modelled and provides a solution for combating repetitive texturing artifacts. We also demonstrate that NeRF textures naturally facilitate continuous level‐of‐detail rendering. Our approach unites the versatility and modelling power of neural networks with the artistic control needed for precise modelling of virtual scenes. While all our training data are currently synthetic, our work provides a recipe that can be further extended to extract complex, hard‐to‐model appearances from real images.

We present a neural scene graph---a modular and controllable representation of scenes with elements that are learned from data. We focus on the forward rendering problem, where the scene graph is provided by the user and references learned elements. The elements correspond to geometry and material definitions of scene objects and constitute the leaves of the graph; we store them as high-dimensional vectors. The position and appearance of scene objects can be adjusted in an artist-friendly manner via familiar transformations, e.g. translation, bending, or color hue shift, which are stored in the inner nodes of the graph. In order to apply a (non-linear) transformation to a learned vector, we adopt the concept of linearizing a problem by lifting it into higher dimensions: we first encode the transformation into a high-dimensional matrix and then apply it by standard matrix-vector multiplication. The transformations are encoded using neural networks. We render the scene graph using a streaming neural renderer, which can handle graphs with a varying number of objects, and thereby facilitates scalability. Our results demonstrate a precise control over the learned object representations in a number of animated 2D and 3D scenes. Despite the limited visual complexity, our work presents a step towards marrying traditional editing mechanisms with learned representations, and towards high-quality, controllable neural rendering.

We present a real-time neural radiance caching method for path-traced global illumination. Our system is designed to handle fully dynamic scenes, and makes no assumptions about the lighting, geometry, and materials. The data-driven nature of our approach sidesteps many difficulties of caching algorithms, such as locating, interpolating, and updating cache points. Since pretraining neural networks to handle novel, dynamic scenes is a formidable generalization challenge, we do away with pretraining and instead achieve generalization via adaptation , i.e. we opt for training the radiance cache while rendering. We employ self-training to provide low-noise training targets and simulate infinite-bounce transport by merely iterating few-bounce training updates. The updates and cache queries incur a mild overhead---about 2.6ms on full HD resolution---thanks to a streaming implementation of the neural network that fully exploits modern hardware. We demonstrate significant noise reduction at the cost of little induced bias, and report state-of-the-art, real-time performance on a number of challenging scenarios.

We present a technique for adaptively partitioning neural scene representations. Our method disentangles lighting, material, and geometric information yielding a scene representation that preserves the orthogonality of these components, improves interpretability of the model, and allows compositing new scenes by mixing components of existing ones. The proposed adaptive partitioning respects the uneven entropy of individual components and permits compressing the scene representation to lower its memory footprint and potentially reduce the evaluation cost of the model. Furthermore, the partitioned representation enables an in-depth analysis of existing image generators. We compare the flow of information through individual partitions, and by contrasting it to the impact of additional inputs (G-buffer), we are able to identify the roots of undesired visual artifacts, and propose one possible solution to remedy the poor performance. We also demonstrate the benefits of complementing traditional forward renderers by neural representations and synthesis, e.g. to infer expensive shading effects, and show how these could improve production rendering in the future if developed further.

We propose neural control variates (NCV) for unbiased variance reduction in parametric Monte Carlo integration. So far, the core challenge of applying the method of control variates has been finding a good approximation of the integrand that is cheap to integrate. We show that a set of neural networks can face that challenge: a normalizing flow that approximates the shape of the integrand and another neural network that infers the solution of the integral equation. We also propose to leverage a neural importance sampler to estimate the difference between the original integrand and the learned control variate. To optimize the resulting parametric estimator, we derive a theoretically optimal, variance-minimizing loss function, and propose an alternative, composite loss for stable online training in practice. When applied to light transport simulation, neural control variates are capable of matching the state-of-the-art performance of other unbiased approaches, while providing means to develop more performant, practical solutions. Specifically, we show that the learned light-field approximation is of sufficient quality for high-order bounces, allowing us to omit the error correction and thereby dramatically reduce the noise at the cost of negligible visible bias.

Abstract We present a novel algorithm to denoise deep Monte Carlo renderings, in which pixels contain multiple colour values, each for a different range of depths. Deep images are a more expressive representation of the scene than conventional flat images. However, since each depth bin receives only a fraction of the flat pixel's samples, denoising the bins is harder due to the less accurate mean and variance estimates. Furthermore, deep images lack a regular structure in depth—the number of depth bins and their depth ranges vary across pixels. This prevents a straightforward application of patch‐based distance metrics frequently used to improve the robustness of existing denoising filters. We address these constraints by combining a flat image‐space non‐local means filter operating on pixel colours with a deep cross‐bilateral filter operating on auxiliary features (albedo, normal, etc.). Our approach significantly reduces noise in deep images while preserving their structure. To our best knowledge, our algorithm is the first to enable efficient deep‐compositing workflows with denoised Monte Carlo renderings. We demonstrate the performance of our filter on a range of scenes highlighting the challenges and advantages of denoising deep images.

We propose to use deep neural networks for generating samples in Monte Carlo integration. Our work is based on non-linear independent components estimation (NICE), which we extend in numerous ways to improve performance and enable its application to integration problems. First, we introduce piecewise-polynomial coupling transforms that greatly increase the modeling power of individual coupling layers. Second, we propose to preprocess the inputs of neural networks using one-blob encoding, which stimulates localization of computation and improves inference. Third, we derive a gradient-descent-based optimization for the Kullback-Leibler and the χ 2 divergence for the specific application of Monte Carlo integration with unnormalized stochastic estimates of the target distribution. Our approach enables fast and accurate inference and efficient sample generation independently of the dimensionality of the integration domain. We show its benefits on generating natural images and in two applications to light-transport simulation: first, we demonstrate learning of joint path-sampling densities in the primary sample space and importance sampling of multi-dimensional path prefixes thereof. Second, we use our technique to extract conditional directional densities driven by the product of incident illumination and the BSDF in the rendering equation, and we leverage the densities for path guiding. In all applications, our approach yields on-par or higher performance than competing techniques at equal sample count.

We present a modular convolutional architecture for denoising rendered images. We expand on the capabilities of kernel-predicting networks by combining them with a number of task-specific modules, and optimizing the assembly using an asymmetric loss. The source-aware encoder---the first module in the assembly---extracts low-level features and embeds them into a common feature space, enabling quick adaptation of a trained network to novel data. The spatial and temporal modules extract abstract, high-level features for kernel-based reconstruction, which is performed at three different spatial scales to reduce low-frequency artifacts. The complete network is trained using a class of asymmetric loss functions that are designed to preserve details and provide the user with a direct control over the variance-bias trade-off during inference. We also propose an error-predicting module for inferring reconstruction error maps that can be used to drive adaptive sampling. Finally, we present a theoretical analysis of convergence rates of kernel-predicting architectures, shedding light on why kernel prediction performs better than synthesizing the colors directly, complementing the empirical evidence presented in this and previous works. We demonstrate that our networks attain results that compare favorably to state-of-the-art methods in terms of detail preservation, low-frequency noise removal, and temporal stability on a variety of production and academic datasets.

Regression-based algorithms have shown to be good at denoising Monte Carlo (MC) renderings by leveraging its inexpensive by-products (e.g., feature buffers). However, when using higher-order models to handle complex cases, these techniques often overfit to noise in the input. For this reason, supervised learning methods have been proposed that train on a large collection of reference examples, but they use explicit filters that limit their denoising ability. To address these problems, we propose a novel, supervised learning approach that allows the filtering kernel to be more complex and general by leveraging a deep convolutional neural network (CNN) architecture. In one embodiment of our framework, the CNN directly predicts the final denoised pixel value as a highly non-linear combination of the input features. In a second approach, we introduce a novel, kernel-prediction network which uses the CNN to estimate the local weighting kernels used to compute each denoised pixel from its neighbors. We train and evaluate our networks on production data and observe improvements over state-of-the-art MC denoisers, showing that our methods generalize well to a variety of scenes. We conclude by analyzing various components of our architecture and identify areas of further research in deep learning for MC denoising.

We address the problem of denoising Monte Carlo renderings by studying existing approaches and proposing a new algorithm that yields state-of-the-art performance on a wide range of scenes. We analyze existing approaches from a theoretical and empirical point of view, relating the strengths and limitations of their corresponding components with an emphasis on production requirements. The observations of our analysis instruct the design of our new filter that offers high-quality results and stable performance. A key observation of our analysis is that using auxiliary buffers (normal, albedo, etc.) to compute the regression weights greatly improves the robustness of zero-order models, but can be detrimental to first-order models. Consequently, our filter performs a first-order regression leveraging a rich set of auxiliary buffers only when fitting the data, and, unlike recent works, considers the pixel color alone when computing the regression weights. We further improve the quality of our output by using a collaborative denoising scheme. Lastly, we introduce a general mean squared error estimator, which can handle the collaborative nature of our filter and its nonlinear weights, to automatically set the bandwidth of our regression kernel.

We explore the theory of integration with control variates in the context of rendering. Our goal is to optimally combine multiple estimators using their covariances. We focus on two applications, re-rendering and gradient-domain rendering, where we exploit coherence between temporally and spatially adjacent pixels. We propose an image-space (iterative) reconstruction scheme that employs control variates to reduce variance. We show that recent works on scene editing and gradient-domain rendering can be directly formulated as control-variate estimators, despite using seemingly different approaches. In particular, we demonstrate the conceptual equivalence of screened Poisson image reconstruction and our iterative reconstruction scheme. Our composite estimators offer practical and simple solutions that improve upon the current state of the art for the two investigated applications.

Abstract Facial appearance capture is now firmly established within academic research and used extensively across various application domains, perhaps most prominently in the entertainment industry through the design of virtual characters in video games and films. While significant progress has occurred over the last two decades, no single survey currently exists that discusses the similarities, differences, and practical considerations of the available appearance capture techniques as applied to human faces. A central difficulty of facial appearance capture is the way light interacts with skin—which has a complex multi‐layered structure—and the interactions that occur below the skin surface can, by definition, only be observed indirectly. In this report, we distinguish between two broad strategies for dealing with this complexity. “Image‐based methods” try to exhaustively capture the exact face appearance under different lighting and viewing conditions, and then render the face through weighted image combinations. “Parametric methods” instead fit the captured reflectance data to some parametric appearance model used during rendering, allowing for a more lightweight and flexible representation but at the cost of potentially increased rendering complexity or inexact reproduction. The goal of this report is to provide an overview that can guide practitioners and researchers in assessing the tradeoffs between current approaches and identifying directions for future advances in facial appearance capture.

Abstract Renderings of animation sequences with physics‐based Monte Carlo light transport simulations are exceedingly costly to generate frame‐by‐frame, yet much of this computation is highly redundant due to the strong coherence in space, time and among samples. A promising approach pursued in prior work entails subsampling the sequence in space, time, and number of samples, followed by image‐based spatio‐temporal upsampling and denoising. These methods can provide significant performance gains, though major issues remain: firstly, in a multiple scattering simulation, the final pixel color is the composite of many different light transport phenomena, and this conflicting information causes artifacts in image‐based methods. Secondly, motion vectors are needed to establish correspondence between the pixels in different frames, but it is unclear how to obtain them for most kinds of light paths (e.g. an object seen through a curved glass panel). To reduce these ambiguities, we propose a general decomposition framework, where the final pixel color is separated into components corresponding to disjoint subsets of the space of light paths. Each component is accompanied by motion vectors and other auxiliary features such as reflectance and surface normals. The motion vectors of specular paths are computed using a temporal extension of manifold exploration and the remaining components use a specialized variant of optical flow. Our experiments show that this decomposition leads to significant improvements in three image‐based applications: denoising, spatial upsampling, and temporal interpolation.

Abstract Monte Carlo integration is firmly established as the basis for most practical realistic image synthesis algorithms because of its flexibility and generality. However, the visual quality of rendered images often suffers from estimator variance, which appears as visually distracting noise. Adaptive sampling and reconstruction algorithms reduce variance by controlling the sampling density and aggregating samples in a reconstruction step, possibly over large image regions. In this paper we survey recent advances in this area. We distinguish between “a priori” methods that analyze the light transport equations and derive sampling rates and reconstruction filters from this analysis, and “a posteriori” methods that apply statistical techniques to sets of samples to drive the adaptive sampling and reconstruction process. They typically estimate the errors of several reconstruction filters, and select the best filter locally to minimize error. We discuss advantages and disadvantages of recent state‐of‐the‐art techniques, and provide visual and quantitative comparisons. Some of these techniques are proving useful in real‐world applications, and we aim to provide an overview for practitioners and researchers to assess these approaches. In addition, we discuss directions for potential further improvements.

We present a generalized framework for gradient-domain Metropolis rendering, and introduce three techniques to reduce sampling artifacts and variance. The first one is a heuristic weighting strategy that combines several sampling techniques to avoid outliers. The second one is an improved mapping to generate offset paths required for computing gradients. Here we leverage the properties of manifold walks in path space to cancel out singularities. Finally, the third technique introduces generalized screen space gradient kernels. This approach aligns the gradient kernels with image structures such as texture edges and geometric discontinuities to obtain sparser gradients than with the conventional gradient kernel. We implement our framework on top of an existing Metropolis sampler, and we demonstrate significant improvements in visual and numerical quality of our results compared to previous work.

Abstract We propose a method that robustly combines color and feature buffers to denoise Monte Carlo renderings. On one hand, feature buffers, such as per pixel normals, textures, or depth, are effective in determining denoising filters because features are highly correlated with rendered images. Filters based solely on features, however, are prone to blurring image details that are not well represented by the features. On the other hand, color buffers represent all details, but they may be less effective to determine filters because they are contaminated by the noise that is supposed to be removed. We propose to obtain filters using a combination of color and feature buffers in an NL‐means and cross‐bilateral filtering framework. We determine a robust weighting of colors and features using a SURE‐based error estimate. We show significant improvements in subjective and quantitative errors compared to the previous state‐of‐theart. We also demonstrate adaptive sampling and space‐time filtering for animations.

We propose a novel approach for image space adaptive sampling and filtering in Monte Carlo rendering. We use an iterative scheme composed of three steps. First, we adaptively distribute samples in the image plane. Second, we denoise the image using a non-linear filter. Third, we estimate the residual per-pixel error of the filtered rendering, and the error estimate guides the sample distribution in the next iteration. The effectiveness of our approach hinges on the use of a state of the art image denoising technique, which we extend to an adaptive rendering framework. A key idea is to split the Monte Carlo samples into two buffers. This improves denoising performance and facilitates variance and error estimation. Our method relies only on the Monte Carlo samples, allowing us to handle arbitrary light transport and lens effects. In addition, it is robust to high noise levels and complex image content. We compare our approach to a state of the art adaptive rendering technique based on adaptive bandwidth selection and demonstrate substantial improvements in terms of both numerical error and visual quality. Our framework is easy to implement on top of standard Monte Carlo renderers and it incurs little computational overhead.

We introduce a novel approach for image space adaptive sampling and reconstruction in Monte Carlo rendering. We greedily minimize relative mean squared error (MSE) by iterating over two steps. First, given a current sample distribution, we optimize over a discrete set of filters at each pixel and select the filter that minimizes the pixel error. Next, given the current filter selection, we distribute additional samples to further reduce MSE. The success of our approach hinges on a robust technique to select suitable per pixel filters. We develop a novel filter selection procedure that robustly solves this problem even with noisy input data. We evaluate our approach using effects such as motion blur, depth of field, interreflections, etc. We provide a comparison to a state-of-the-art algorithm based on wavelet shrinkage and show that we achieve significant improvements in numerical error and visual image quality. Our approach is simple to implement, requires a single user parameter, and is compatible with standard Monte Carlo rendering.