Understanding and modeling lighting effects are fundamental tasks in computer vision and graphics. Classic physically-based rendering (PBR) accurately simulates the light transport, but relies on precise scene representations–explicit 3D geometry, high-quality material properties, and lighting conditions–that are often impractical to obtain in real-world scenarios. Therefore, we introduce DiffusionRenderer, a neural approach that addresses the dual problem of inverse and forward rendering within a holistic framework. Leveraging powerful video diffusion model priors, the inverse rendering model accurately estimates G-buffers from real-world videos, providing an interface for image editing tasks, and training data for the rendering model. Conversely, our rendering model generates photorealistic images from G-buffers without explicit light transport simulation. Specifically, we first train a video diffusion model for inverse rendering on synthetic data, which generalizes well to real-world videos and allows us to auto-label diverse real-world videos. We then co-train our rendering model using both synthetic and auto-labeled real-world data. Experiments demonstrate that DiffusionRenderer effectively approximates inverse and forwards rendering, consistently outperforming the state-of-the-art. Our model enables practical applications from a single video input—including relighting, material editing, and realistic object insertion.

We address the challenge of relighting a single image or video, a task that demands precise scene intrinsic understanding and high-quality light transport synthesis. Existing end-to-end relighting models are often limited by the scarcity of paired multi-illumination data, restricting their ability to generalize across diverse scenes. Conversely, two-stage pipelines that combine inverse and forward rendering can mitigate data requirements but are susceptible to error accumulation and often fail to produce realistic outputs under complex lighting conditions or with sophisticated materials. In this work, we introduce a general-purpose approach that jointly estimates albedo and synthesizes relit outputs in a single pass, harnessing the generative capabilities of video diffusion models. This joint formulation enhances implicit scene comprehension and facilitates the creation of realistic lighting effects and intricate material interactions, such as shadows, reflections, and transparency. Trained on synthetic multi-illumination data and extensive automatically labeled real-world videos, our model demonstrates strong generalization across diverse domains and surpasses previous methods in both visual fidelity and temporal consistency. Our project page is https://research.nvidia.com/labs/toronto-ai/UniRelight/.

Abstract We leverage finetuned video diffusion models, intrinsic decomposition of videos, and physically‐based differentiable rendering to generate high quality materials for 3D models given a text prompt or a single image. We condition a video diffusion model to respect the input geometry and lighting condition. This model produces multiple views of a given 3D model with coherent material properties. Secondly, we use a recent model to extract intrinsics (base color, roughness, metallic) from the generated video. Finally, we use the intrinsics alongside the generated video in a differentiable path tracer to robustly extract PBR materials directly compatible with common content creation tools.

Reconstruction and intrinsic decomposition of scenes from captured imagery would enable many applications such as relighting and virtual object insertion. Recent NeRF based methods achieve impressive fidelity of 3D reconstruction, but bake the lighting and shadows into the radiance field, while mesh-based methods that facilitate intrinsic decomposition through differentiable rendering have not yet scaled to the complexity and scale of outdoor scenes. We present a novel inverse rendering framework for large urban scenes capable of jointly reconstructing the scene geometry, spatially-varying materials, and HDR lighting from a set of posed RGB images with optional depth. Specifically, we use a neural field to account for the primary rays, and use an explicit mesh (reconstructed from the underlying neural field) for modeling secondary rays that produce higher-order lighting effects such as cast shadows. By faithfully disentangling complex geometry and materials from lighting effects, our method enables photorealistic relighting with specular and shadow effects on several outdoor datasets. Moreover, it supports physics-based scene manipulations such as virtual object insertion with ray-traced shadow casting.

We introduce SLANG.D, an extension to the Slang shading language that incorporates first-class automatic differentiation support. The new shading language allows us to transform a Direct3D-based path tracer to be fully differentiable with minor modifications to existing code. SLANG.D enables a shared ecosystem between machine learning frameworks and pre-existing graphics hardware API-based rendering systems, promoting the interchange of components and ideas across these two domains. Our contributions include a differentiable type system designed to ensure type safety and semantic clarity in codebases that blend differentiable and non-differentiable code, language primitives that automatically generate both forward and reverse gradient propagation methods, and a compiler architecture that generates efficient derivative propagation shader code for graphics pipelines. Our compiler supports differentiating code that involves arbitrary control-flow, dynamic dispatch, generics and higher-order differentiation, while providing developers flexible control of checkpointing and gradient aggregation strategies for best performance. Our system allows us to differentiate an existing real-time path tracer, Falcor, with minimal change to its shader code. We show that the compiler-generated derivative kernels perform as efficiently as handwritten ones. In several benchmarks, the SLANG.D code achieves significant speedup when compared to prior automatic differentiation systems.

This work considers gradient-based mesh optimization, where we iteratively optimize for a 3D surface mesh by representing it as the isosurface of a scalar field, an increasingly common paradigm in applications including photogrammetry, generative modeling, and inverse physics. Existing implementations adapt classic isosurface extraction algorithms like Marching Cubes or Dual Contouring; these techniques were designed to extract meshes from fixed, known fields, and in the optimization setting they lack the degrees of freedom to represent high-quality feature-preserving meshes, or suffer from numerical instabilities. We introduce FlexiCubes, an isosurface representation specifically designed for optimizing an unknown mesh with respect to geometric, visual, or even physical objectives. Our main insight is to introduce additional carefully-chosen parameters into the representation, which allow local flexible adjustments to the extracted mesh geometry and connectivity. These parameters are updated along with the underlying scalar field via automatic differentiation when optimizing for a downstream task. We base our extraction scheme on Dual Marching Cubes for improved topological properties, and present extensions to optionally generate tetrahedral and hierarchically-adaptive meshes. Extensive experiments validate FlexiCubes on both synthetic benchmarks and real-world applications, showing that it offers significant improvements in mesh quality and geometric fidelity.

We present an efficient method for joint optimization of topology, materials and lighting from multi-view image observations. Unlike recent multi-view reconstruction approaches, which typically produce entangled 3D representations encoded in neural networks, we output triangle meshes with spatially-varying materials and environment lighting that can be deployed in any traditional graphics engine unmodified. We leverage recent work in differentiable rendering, coordinate-based networks to compactly represent volumetric texturing, alongside differentiable marching tetrahedrons to enable gradient-based optimization directly on the surface mesh. Finally, we introduce a differentiable formulation of the split sum approximation of environment lighting to efficiently recover all-frequency lighting. Experiments show our extracted models used in advanced scene editing, material decomposition, and high quality view interpolation, all running at interactive rates in triangle-based renderers (rasterizers and path tracers).

Recent advances in differentiable rendering have enabled high-quality reconstruction of 3D scenes from multi-view images. Most methods rely on simple rendering algorithms: pre-filtered direct lighting or learned representations of irradiance. We show that a more realistic shading model, incorporating ray tracing and Monte Carlo integration, substantially improves decomposition into shape, materials & lighting. Unfortunately, Monte Carlo integration provides estimates with significant noise, even at large sample counts, which makes gradient-based inverse rendering very challenging. To address this, we incorporate multiple importance sampling and denoising in a novel inverse rendering pipeline. This improves convergence and enables gradient-based optimization at low sample counts. We present an efficient method to jointly reconstruct geometry (explicit triangle meshes), materials, and lighting, which substantially improves material and light separation compared to previous work. We argue that denoising can become an integral part of high quality inverse rendering pipelines.

Abstract Despite recent advances in Monte Carlo path tracing at interactive rates, denoised image sequences generated with few samples per‐pixel often yield temporally unstable results and loss of high‐frequency details. We present a novel adaptive rendering method that increases temporal stability and image fidelity of low sample count path tracing by distributing samples via spatio‐temporal joint optimization of sampling and denoising. Adding temporal optimization to the sample predictor enables it to learn spatio‐temporal sampling strategies such as placing more samples in disoccluded regions, tracking specular highlights, etc; adding temporal feedback to the denoiser boosts the effective input sample count and increases temporal stability. The temporal approach also allows us to remove the initial uniform sampling step typically present in adaptive sampling algorithms. The sample predictor and denoiser are deep neural networks that we co‐train end‐to‐end over multiple consecutive frames. Our approach is scalable, allowing trade‐off between quality and performance, and runs at near real‐time rates while achieving significantly better image quality and temporal stability than previous methods.

Abstract We propose a novel approach for denoising Monte Carlo path traced images, which uses data from individual samples rather than relying on pixel aggregates. Samples are partitioned into layers, which are filtered separately, giving the network more freedom to handle outliers and complex visibility. Finally the layers are composited front‐to‐back using alpha blending. The system is trained end‐to‐end, with learned layer partitioning, filter kernels, and compositing. We obtain similar image quality as recent state‐of‐the‐art sample based denoisers at a fraction of the computational cost and memory requirements.

We apply basic statistical reasoning to signal reconstruction by machine learning - learning to map corrupted observations to clean signals - with a simple and powerful conclusion: it is possible to learn to restore images by only looking at corrupted examples, at performance at and sometimes exceeding training using clean data, without explicit image priors or likelihood models of the corruption. In practice, we show that a single model learns photographic noise removal, denoising synthetic Monte Carlo images, and reconstruction of undersampled MRI scans - all corrupted by different processes - based on noisy data only.

We present a texture space caching and reconstruction system for Monte Carlo ray tracing. Our system gathers and filters shading on-demand, including querying secondary rays, directly within a filter footprint around the current shading point. We shade on local grids in texture space with primary visibility decoupled from shading. Unique filters can be applied per material, where any terms of the shader can be chosen to be included in each kernel. This is a departure from recent screen space image reconstruction techniques, which typically use a single, complex kernel with a set of large auxiliary guide images as input. We show a number of high-performance use cases for our system, including interactive denoising of Monte Carlo ray tracing with motion/defocus blur, spatial and temporal shading reuse, cached product importance sampling, and filters based on linear regression in texture space.

Abstract Given a stochastic shadow map rendered with motion blur, our goal is to render an image from the eye with motion‐blurred shadows with as little noise as possible. We use a layered approach in the shadow map and reproject samples along the average motion vector, and then perform lookups in this representation. Our results include substantially improved shadow quality compared to previous work and a fast graphics processing unit (GPU) implementation. In addition, we devise a set of scenes that are designed to bring out and show problematic cases for motion‐blurred shadows. These scenes have difficult occlusion characteristics, and may be used in future research on this topic.

We present a novel algorithm for reconstructing high-quality defocus blur from a sparsely sampled light field. Our algorithm builds upon recent developments in the area of sheared reconstruction filters and significantly improves reconstruction quality and performance. While previous filtering techniques can be ineffective in regions with complex occlusion, our algorithm handles such scenarios well by partitioning the input samples into depth layers. These depth layers are filtered independently and then combined together, taking into account inter-layer visibility. We also introduce a new separable formulation of sheared reconstruction filters that achieves real-time preformance on a modern GPU and is more than two orders of magnitude faster than previously published techniques.

Abstract Light field reconstruction algorithms can substantially decrease the noise in stochastically rendered images. Recent algorithms for defocus blur alone are both fast and accurate. However, motion blur is a considerably more complex type of camera effect, and as a consequence, current algorithms are either slow or too imprecise to use in high quality rendering. We extend previous work on real‐time light field reconstruction for defocus blur to handle the case of simultaneous defocus and motion blur. By carefully introducing a few approximations, we derive a very efficient sheared reconstruction filter, which produces high quality images even for a low number of input samples. Our algorithm is temporally robust, and is about two orders of magnitude faster than previous work, making it suitable for both real‐time rendering and as a post‐processing pass for offline rendering.

Real-time rendering with true motion and defocus blur remains an elusive goal for application developers. In recent years, substantial progress has been made in the areas of rasterization, shading, and reconstruction for stochastic rendering. However, we have yet to see an efficient method for decoupled sampling that can be implemented on current or near-future graphics processors. In this paper, we propose one such algorithm that leverages the capability of modern GPUs to perform unordered memory accesses from within shaders. Our algorithm builds per-pixel primitive lists in canonical shading space. All shading then takes place in a single, non-multisampled forward rendering pass using conservative rasterization. This pass exploits the rasterization and shading hardware to perform shading very efficiently, and only samples that are visible in the final image are shaded. Last, the shading samples are gathered and filtered to create the final image. The input to our algorithm can be generated using a variety of methods, of which we show examples of interactive stochastic and interleaved rasterization, as well as ray tracing.

Abstract In this paper, we derive compact representations of the depth function for a triangle undergoing motion or defocus blur. Unlike a static primitive, where the depth function is planar, the depth function is a rational function in time and the lens parameters. Furthermore, we show how these compact depth functions can be used to design an efficient depth buffer compressor/decompressor, which significantly lowers total depth buffer bandwidth usage for a range of test scenes. In addition, our compressor/decompressor is simpler in the number of operations needed to execute, which makes our algorithm more amenable for hardware implementation than previous methods.

Stochastic sampling in time and over the lens is essential to produce photo-realistic images, and it has the potential to revolutionize real-time graphics. In this paper, we take an architectural view of the problem and propose a novel hardware architecture for efficient shading in the context of stochastic rendering. We replace previous caching mechanisms by a sorting step to extract coherence, thereby ensuring that only non-occluded samples are shaded. The memory bandwidth is kept at a minimum by operating on tiles and using new buffer compression methods. Our architecture has several unique benefits not traditionally associated with deferred shading. First, shading is performed in primitive order, which enables late shading of vertex attributes and avoids the need to generate a G-buffer of pre-interpolated vertex attributes. Second, we support state changes, e.g., change of shaders and resources in the deferred shading pass, avoiding the need for a single über-shader. We perform an extensive architectural simulation to quantify the benefits of our algorithm on real workloads.

Abstract For depth of field (DOF) rasterization, it is often desired to have an efficient tile versus triangle test, which can conservatively compute which samples on the lens that need to execute the sample‐in‐triangle test. We present a novel test for this, which is optimal in the sense that the region on the lens cannot be further reduced. Our test is based on removing half‐space regions of the (u, v) ‐space on the lens, from where the triangle definitely cannot be seen through a tile of pixels. We find the intersection of all such regions exactly, and the resulting region can be used to reduce the number of sample‐in‐triangle tests that need to be performed. Our main contribution is that the theory we develop provides a limit for how efficient a practical tile versus defocused triangle test ever can become. To verify our work, we also develop a conceptual implementation for DOF rasterization based on our new theory. We show that the number of arithmetic operations involved in the rasterization process can be reduced. More importantly, with a tile test, multi‐sampling anti‐aliasing can be used which may reduce shader executions and the related memory bandwidth usage substantially. In general, this can be translated to a performance increase and/or power savings.

Abstract We present user‐controllable and plausible defocus blur for a stochastic rasterizer. We modify circle of confusion coefficients per vertex to express more general defocus blur, and show how the method can be applied to limit the foreground blur, extend the in‐focus range, simulate tilt‐shift photography, and specify per‐object defocus blur. Furthermore, with two simplifying assumptions, we show that existing triangle coverage tests and tile culling tests can be used with very modest modifications. Our solution is temporally stable and handles simultaneous motion blur and depth of field.

Graphics processing units supporting tessellation of curved surfaces with displacement mapping exist today. Still, to our knowledge, culling only occurs after tessellation, that is, after the base primitives have been tessellated into triangles. We introduce an algorithm for automatically computing tight positional and normal bounds on the fly for a base primitive. These bounds are derived from an arbitrary vertex shader program, which may include a curved surface evaluation and different types of displacements, for example. The obtained bounds are used for backface, view frustum, and occlusion culling before tessellation. For highly tessellated scenes, we show that up to 80% of the vertex shader instructions can be avoided, which implies an “instruction speedup” of 5×. Our technique can also be used for offline software rendering.

Abstract The use of high dynamic range (HDR) textures in real‐time graphics applications can increase realism and provide a more vivid experience. However, the increased bandwidth and storage requirements for uncompressed HDR data can become a major bottleneck. Hence, several recent algorithms for HDR texture compression have been proposed. In this paper, we discuss several practical issues one has to confront in order to develop and implement HDR texture compression schemes. These include improved texture filtering and efficient offline compression. For compression, we describe how Procrustes analysis can be used to quickly match a predefined template shape against chrominance data. To reduce the cost of HDR texture filtering, we perform filtering prior to the colour transformation, and use a simple trick to reduce the incurred errors. We also introduce a number of novel compression modes, which can be combined with existing compression schemes, or used on their own.

In this paper, we break new ground by presenting algorithms for fixed-rate compression of high dynamic range textures at low bit rates. First, the S3TC low dynamic range texture compression scheme is extended in order to enable compression of HDR data. Second, we introduce a novel robust algorithm that offers superior image quality. Our algorithm can be efficiently implemented in hardware, and supports textures with a dynamic range of over 10 9 :1. At a fixed rate of 8 bits per pixel, we obtain results virtually indistinguishable from uncompressed HDR textures at 48 bits per pixel. Our research can have a big impact on graphics hardware and real-time rendering, since HDR texturing suddenly becomes affordable.