Ben Mildenhall

Ben Poole, Ajay Jain, Jonathan T. Barron et al. · berkeley

Recent breakthroughs in text-to-image synthesis have been driven by diffusion models trained on billions of image-text pairs. Adapting this approach to 3D synthesis would require large-scale datasets of labeled 3D data and efficient architectures for denoising 3D data, neither of which currently exist. In this work, we circumvent these limitations by using a pretrained 2D text-to-image diffusion model to perform text-to-3D synthesis. We introduce a loss based on probability density distillation that enables the use of a 2D diffusion model as a prior for optimization of a parametric image generator. Using this loss in a DeepDream-like procedure, we optimize a randomly-initialized 3D model (a Neural Radiance Field, or NeRF) via gradient descent such that its 2D renderings from random angles achieve a low loss. The resulting 3D model of the given text can be viewed from any angle, relit by arbitrary illumination, or composited into any 3D environment. Our approach requires no 3D training data and no modifications to the image diffusion model, demonstrating the effectiveness of pretrained image diffusion models as priors.

Lior Yariv, Peter Hedman, Christian Reiser et al.

We present a method for reconstructing high-quality meshes of large unbounded real-world scenes suitable for photorealistic novel view synthesis. We first optimize a hybrid neural volume-surface scene representation designed to have well-behaved level sets that correspond to surfaces in the scene. We then bake this representation into a high-quality triangle mesh, which we equip with a simple and fast view-dependent appearance model based on spherical Gaussians. Finally, we optimize this baked representation to best reproduce the captured viewpoints, resulting in a model that can leverage accelerated polygon rasterization pipelines for real-time view synthesis on commodity hardware. Our approach outperforms previous scene representations for real-time rendering in terms of accuracy, speed, and power consumption, and produces high quality meshes that enable applications such as appearance editing and physical simulation.

Christian Reiser, Richard Szeliski, Dor Verbin et al.

Neural radiance fields enable state-of-the-art photorealistic view synthesis. However, existing radiance field representations are either too compute-intensive for real-time rendering or require too much memory to scale to large scenes. We present a Memory-Efficient Radiance Field (MERF) representation that achieves real-time rendering of large-scale scenes in a browser. MERF reduces the memory consumption of prior sparse volumetric radiance fields using a combination of a sparse feature grid and high-resolution 2D feature planes. To support large-scale unbounded scenes, we introduce a novel contraction function that maps scene coordinates into a bounded volume while still allowing for efficient ray-box intersection. We design a lossless procedure for baking the parameterization used during training into a model that achieves real-time rendering while still preserving the photorealistic view synthesis quality of a volumetric radiance field.

Jonathan T. Barron, Ben Mildenhall, Dor Verbin et al.

Neural Radiance Field training can be accelerated through the use of grid-based representations in NeRF's learned mapping from spatial coordinates to colors and volumetric density. However, these grid-based approaches lack an explicit understanding of scale and therefore often introduce aliasing, usually in the form of jaggies or missing scene content. Anti-aliasing has previously been addressed by mip-NeRF 360, which reasons about sub-volumes along a cone rather than points along a ray, but this approach is not natively compatible with current grid-based techniques. We show how ideas from rendering and signal processing can be used to construct a technique that combines mip-NeRF 360 and grid-based models such as Instant NGP to yield error rates that are 8% - 77% lower than either prior technique, and that trains 24x faster than mip-NeRF 360.

Ryan Po, Wang Yifan, Vladislav Golyanik et al.

The field of visual computing is rapidly advancing due to the emergence of generative artificial intelligence (AI), which unlocks unprecedented capabilities for the generation, editing, and reconstruction of images, videos, and 3D scenes. In these domains, diffusion models are the generative AI architecture of choice. Within the last year alone, the literature on diffusion-based tools and applications has seen exponential growth and relevant papers are published across the computer graphics, computer vision, and AI communities with new works appearing daily on arXiv. This rapid growth of the field makes it difficult to keep up with all recent developments. The goal of this state-of-the-art report (STAR) is to introduce the basic mathematical concepts of diffusion models, implementation details and design choices of the popular Stable Diffusion model, as well as overview important aspects of these generative AI tools, including personalization, conditioning, inversion, among others. Moreover, we give a comprehensive overview of the rapidly growing literature on diffusion-based generation and editing, categorized by the type of generated medium, including 2D images, videos, 3D objects, locomotion, and 4D scenes. Finally, we discuss available datasets, metrics, open challenges, and social implications. This STAR provides an intuitive starting point to explore this exciting topic for researchers, artists, and practitioners alike.

Amit Raj, Srinivas Kaza, Ben Poole et al.

We present DreamBooth3D, an approach to personalize text-to-3D generative models from as few as 3-6 casually captured images of a subject. Our approach combines recent advances in personalizing text-to-image models (DreamBooth) with text-to-3D generation (DreamFusion). We find that naively combining these methods fails to yield satisfactory subject-specific 3D assets due to personalized text-to-image models overfitting to the input viewpoints of the subject. We overcome this through a 3-stage optimization strategy where we jointly leverage the 3D consistency of neural radiance fields together with the personalization capability of text-to-image models. Our method can produce high-quality, subject-specific 3D assets with text-driven modifications such as novel poses, colors and attributes that are not seen in any of the input images of the subject.

Keunhong Park, Philipp Henzler, Ben Mildenhall et al.

Neural Radiance Fields (NeRF) can be optimized to obtain high-fidelity 3D scene reconstructions of objects and large-scale scenes. However, NeRFs require accurate camera parameters as input -- inaccurate camera parameters result in blurry renderings. Extrinsic and intrinsic camera parameters are usually estimated using Structure-from-Motion (SfM) methods as a pre-processing step to NeRF, but these techniques rarely yield perfect estimates. Thus, prior works have proposed jointly optimizing camera parameters alongside a NeRF, but these methods are prone to local minima in challenging settings. In this work, we analyze how different camera parameterizations affect this joint optimization problem, and observe that standard parameterizations exhibit large differences in magnitude with respect to small perturbations, which can lead to an ill-conditioned optimization problem. We propose using a proxy problem to compute a whitening transform that eliminates the correlation between camera parameters and normalizes their effects, and we propose to use this transform as a preconditioner for the camera parameters during joint optimization. Our preconditioned camera optimization significantly improves reconstruction quality on scenes from the Mip-NeRF 360 dataset: we reduce error rates (RMSE) by 67% compared to state-of-the-art NeRF approaches that do not optimize for cameras like Zip-NeRF, and by 29% relative to state-of-the-art joint optimization approaches using the camera parameterization of SCNeRF. Our approach is easy to implement, does not significantly increase runtime, can be applied to a wide variety of camera parameterizations, and can straightforwardly be incorporated into other NeRF-like models.

Yifan Jiang, Peter Hedman, Ben Mildenhall et al.

Neural Radiance Fields (NeRFs) are a powerful representation for modeling a 3D scene as a continuous function. Though NeRF is able to render complex 3D scenes with view-dependent effects, few efforts have been devoted to exploring its limits in a high-resolution setting. Specifically, existing NeRF-based methods face several limitations when reconstructing high-resolution real scenes, including a very large number of parameters, misaligned input data, and overly smooth details. In this work, we conduct the first pilot study on training NeRF with high-resolution data and propose the corresponding solutions: 1) marrying the multilayer perceptron (MLP) with convolutional layers which can encode more neighborhood information while reducing the total number of parameters; 2) a novel training strategy to address misalignment caused by moving objects or small camera calibration errors; and 3) a high-frequency aware loss. Our approach is nearly free without introducing obvious training/testing costs, while experiments on different datasets demonstrate that it can recover more high-frequency details compared with the current state-of-the-art NeRF models. Project page: \url{https://yifanjiang.net/alignerf.}

Andrea Tagliasacchi, Ben Mildenhall

Neural Radiance Fields employ simple volume rendering as a way to overcome the challenges of differentiating through ray-triangle intersections by leveraging a probabilistic notion of visibility. This is achieved by assuming the scene is composed by a cloud of light-emitting particles whose density changes in space. This technical report summarizes the derivations for differentiable volume rendering. It is a condensed version of previous reports, but rewritten in the context of NeRF, and adopting its commonly used notation.

Benjamin Attal, Dor Verbin, Ben Mildenhall et al.

State-of-the-art techniques for 3D reconstruction are largely based on volumetric scene representations, which require sampling multiple points to compute the color arriving along a ray. Using these representations for more general inverse rendering -- reconstructing geometry, materials, and lighting from observed images -- is challenging because recursively path-tracing such volumetric representations is expensive. Recent works alleviate this issue through the use of radiance caches: data structures that store the steady-state, infinite-bounce radiance arriving at any point from any direction. However, these solutions rely on approximations that introduce bias into the renderings and, more importantly, into the gradients used for optimization. We present a method that avoids these approximations while remaining computationally efficient. In particular, we leverage two techniques to reduce variance for unbiased estimators of the rendering equation: (1) an occlusion-aware importance sampler for incoming illumination and (2) a fast cache architecture that can be used as a control variate for the radiance from a high-quality, but more expensive, volumetric cache. We show that by removing these biases our approach improves the generality of radiance cache based inverse rendering, as well as increasing quality in the presence of challenging light transport effects such as specular reflections.

Rundi Wu, Ben Mildenhall, Philipp Henzler et al.

3D reconstruction methods such as Neural Radiance Fields (NeRFs) excel at rendering photorealistic novel views of complex scenes. However, recovering a high-quality NeRF typically requires tens to hundreds of input images, resulting in a time-consuming capture process. We present ReconFusion to reconstruct real-world scenes using only a few photos. Our approach leverages a diffusion prior for novel view synthesis, trained on synthetic and multiview datasets, which regularizes a NeRF-based 3D reconstruction pipeline at novel camera poses beyond those captured by the set of input images. Our method synthesizes realistic geometry and texture in underconstrained regions while preserving the appearance of observed regions. We perform an extensive evaluation across various real-world datasets, including forward-facing and 360-degree scenes, demonstrating significant performance improvements over previous few-view NeRF reconstruction approaches.

Christian Reiser, Stephan Garbin, Pratul P. Srinivasan et al.

While surface-based view synthesis algorithms are appealing due to their low computational requirements, they often struggle to reproduce thin structures. In contrast, more expensive methods that model the scene's geometry as a volumetric density field (e.g. NeRF) excel at reconstructing fine geometric detail. However, density fields often represent geometry in a "fuzzy" manner, which hinders exact localization of the surface. In this work, we modify density fields to encourage them to converge towards surfaces, without compromising their ability to reconstruct thin structures. First, we employ a discrete opacity grid representation instead of a continuous density field, which allows opacity values to discontinuously transition from zero to one at the surface. Second, we anti-alias by casting multiple rays per pixel, which allows occlusion boundaries and subpixel structures to be modelled without using semi-transparent voxels. Third, we minimize the binary entropy of the opacity values, which facilitates the extraction of surface geometry by encouraging opacity values to binarize towards the end of training. Lastly, we develop a fusion-based meshing strategy followed by mesh simplification and appearance model fitting. The compact meshes produced by our model can be rendered in real-time on mobile devices and achieve significantly higher view synthesis quality compared to existing mesh-based approaches.

Dave Epstein, Ben Poole, Ben Mildenhall et al. · berkeley

We introduce a method to generate 3D scenes that are disentangled into their component objects. This disentanglement is unsupervised, relying only on the knowledge of a large pretrained text-to-image model. Our key insight is that objects can be discovered by finding parts of a 3D scene that, when rearranged spatially, still produce valid configurations of the same scene. Concretely, our method jointly optimizes multiple NeRFs from scratch - each representing its own object - along with a set of layouts that composite these objects into scenes. We then encourage these composited scenes to be in-distribution according to the image generator. We show that despite its simplicity, our approach successfully generates 3D scenes decomposed into individual objects, enabling new capabilities in text-to-3D content creation. For results and an interactive demo, see our project page at https://dave.ml/layoutlearning/

Pratul P. Srinivasan, Stephan J. Garbin, Dor Verbin et al.

Existing UV mapping algorithms are designed to operate on well-behaved meshes, instead of the geometry representations produced by state-of-the-art 3D reconstruction and generation techniques. As such, applying these methods to the volume densities recovered by neural radiance fields and related techniques (or meshes triangulated from such fields) results in texture atlases that are too fragmented to be useful for tasks such as view synthesis or appearance editing. We present a UV mapping method designed to operate on geometry produced by 3D reconstruction and generation techniques. Instead of computing a mapping defined on a mesh's vertices, our method Nuvo uses a neural field to represent a continuous UV mapping, and optimizes it to be a valid and well-behaved mapping for just the set of visible points, i.e. only points that affect the scene's appearance. We show that our model is robust to the challenges posed by ill-behaved geometry, and that it produces editable UV mappings that can represent detailed appearance.

Dor Verbin, Pratul P. Srinivasan, Peter Hedman et al.

Neural Radiance Fields (NeRFs) typically struggle to reconstruct and render highly specular objects, whose appearance varies quickly with changes in viewpoint. Recent works have improved NeRF's ability to render detailed specular appearance of distant environment illumination, but are unable to synthesize consistent reflections of closer content. Moreover, these techniques rely on large computationally-expensive neural networks to model outgoing radiance, which severely limits optimization and rendering speed. We address these issues with an approach based on ray tracing: instead of querying an expensive neural network for the outgoing view-dependent radiance at points along each camera ray, our model casts reflection rays from these points and traces them through the NeRF representation to render feature vectors which are decoded into color using a small inexpensive network. We demonstrate that our model outperforms prior methods for view synthesis of scenes containing shiny objects, and that it is the only existing NeRF method that can synthesize photorealistic specular appearance and reflections in real-world scenes, while requiring comparable optimization time to current state-of-the-art view synthesis models.

Xiaojuan Wang, Janne Kontkanen, Brian Curless et al. · uw

We present a method that uses a text-to-image model to generate consistent content across multiple image scales, enabling extreme semantic zooms into a scene, e.g., ranging from a wide-angle landscape view of a forest to a macro shot of an insect sitting on one of the tree branches. We achieve this through a joint multi-scale diffusion sampling approach that encourages consistency across different scales while preserving the integrity of each individual sampling process. Since each generated scale is guided by a different text prompt, our method enables deeper levels of zoom than traditional super-resolution methods that may struggle to create new contextual structure at vastly different scales. We compare our method qualitatively with alternative techniques in image super-resolution and outpainting, and show that our method is most effective at generating consistent multi-scale content.

Quankai Gao, Qiangeng Xu, Zhe Cao et al.

Creating 4D fields of Gaussian Splatting from images or videos is a challenging task due to its under-constrained nature. While the optimization can draw photometric reference from the input videos or be regulated by generative models, directly supervising Gaussian motions remains underexplored. In this paper, we introduce a novel concept, Gaussian flow, which connects the dynamics of 3D Gaussians and pixel velocities between consecutive frames. The Gaussian flow can be efficiently obtained by splatting Gaussian dynamics into the image space. This differentiable process enables direct dynamic supervision from optical flow. Our method significantly benefits 4D dynamic content generation and 4D novel view synthesis with Gaussian Splatting, especially for contents with rich motions that are hard to be handled by existing methods. The common color drifting issue that happens in 4D generation is also resolved with improved Guassian dynamics. Superior visual quality on extensive experiments demonstrates our method's effectiveness. Quantitative and qualitative evaluations show that our method achieves state-of-the-art results on both tasks of 4D generation and 4D novel view synthesis. Project page: https://zerg-overmind.github.io/GaussianFlow.github.io/

Dor Verbin, Ben Mildenhall, Peter Hedman et al.

Decomposing an object's appearance into representations of its materials and the surrounding illumination is difficult, even when the object's 3D shape is known beforehand. This problem is especially challenging for diffuse objects: it is ill-conditioned because diffuse materials severely blur incoming light, and it is ill-posed because diffuse materials under high-frequency lighting can be indistinguishable from shiny materials under low-frequency lighting. We show that it is possible to recover precise materials and illumination -- even from diffuse objects -- by exploiting unintended shadows, like the ones cast onto an object by the photographer who moves around it. These shadows are a nuisance in most previous inverse rendering pipelines, but here we exploit them as signals that improve conditioning and help resolve material-lighting ambiguities. We present a method based on differentiable Monte Carlo ray tracing that uses images of an object to jointly recover its spatially-varying materials, the surrounding illumination environment, and the shapes of the unseen light occluders who inadvertently cast shadows upon it.

Matthew Tancik, Vincent Casser, Xinchen Yan et al.

We present Block-NeRF, a variant of Neural Radiance Fields that can represent large-scale environments. Specifically, we demonstrate that when scaling NeRF to render city-scale scenes spanning multiple blocks, it is vital to decompose the scene into individually trained NeRFs. This decomposition decouples rendering time from scene size, enables rendering to scale to arbitrarily large environments, and allows per-block updates of the environment. We adopt several architectural changes to make NeRF robust to data captured over months under different environmental conditions. We add appearance embeddings, learned pose refinement, and controllable exposure to each individual NeRF, and introduce a procedure for aligning appearance between adjacent NeRFs so that they can be seamlessly combined. We build a grid of Block-NeRFs from 2.8 million images to create the largest neural scene representation to date, capable of rendering an entire neighborhood of San Francisco.

Yifan Jiang, Bartlomiej Wronski, Ben Mildenhall et al.

Most image denoising networks apply a single set of static convolutional kernels across the entire input image. This is sub-optimal for natural images, as they often consist of heterogeneous visual patterns. Dynamic convolution tries to address this issue by using per-pixel convolution kernels, but this greatly increases computational cost. In this work, we present Malleable Convolution (MalleConv), which performs spatial-varying processing with minimal computational overhead. MalleConv uses a smaller set of spatially-varying convolution kernels, a compromise between static and per-pixel convolution kernels. These spatially-varying kernels are produced by an efficient predictor network running on a downsampled input, making them much more efficient to compute than per-pixel kernels produced by a full-resolution image, and also enlarging the network's receptive field compared with static kernels. These kernels are then jointly upsampled and applied to a full-resolution feature map through an efficient on-the-fly slicing operator with minimum memory overhead. To demonstrate the effectiveness of MalleConv, we use it to build an efficient denoising network we call MalleNet. MalleNet achieves high-quality results without very deep architectures, making it 8.9x faster than the best performing denoising algorithms while achieving similar visual quality. We also show that a single MalleConv layer added to a standard convolution-based backbone can significantly reduce the computational cost or boost image quality at a similar cost. More information is on our project page: \url{https://yifanjiang.net/MalleConv.html}

Dor Verbin, Peter Hedman, Ben Mildenhall et al.

Neural Radiance Fields (NeRF) is a popular view synthesis technique that represents a scene as a continuous volumetric function, parameterized by multilayer perceptrons that provide the volume density and view-dependent emitted radiance at each location. While NeRF-based techniques excel at representing fine geometric structures with smoothly varying view-dependent appearance, they often fail to accurately capture and reproduce the appearance of glossy surfaces. We address this limitation by introducing Ref-NeRF, which replaces NeRF's parameterization of view-dependent outgoing radiance with a representation of reflected radiance and structures this function using a collection of spatially-varying scene properties. We show that together with a regularizer on normal vectors, our model significantly improves the realism and accuracy of specular reflections. Furthermore, we show that our model's internal representation of outgoing radiance is interpretable and useful for scene editing.

Barbara Roessle, Jonathan T. Barron, Ben Mildenhall et al.

Neural radiance fields (NeRF) encode a scene into a neural representation that enables photo-realistic rendering of novel views. However, a successful reconstruction from RGB images requires a large number of input views taken under static conditions - typically up to a few hundred images for room-size scenes. Our method aims to synthesize novel views of whole rooms from an order of magnitude fewer images. To this end, we leverage dense depth priors in order to constrain the NeRF optimization. First, we take advantage of the sparse depth data that is freely available from the structure from motion (SfM) preprocessing step used to estimate camera poses. Second, we use depth completion to convert these sparse points into dense depth maps and uncertainty estimates, which are used to guide NeRF optimization. Our method enables data-efficient novel view synthesis on challenging indoor scenes, using as few as 18 images for an entire scene.

Ajay Jain, Ben Mildenhall, Jonathan T. Barron et al.

We combine neural rendering with multi-modal image and text representations to synthesize diverse 3D objects solely from natural language descriptions. Our method, Dream Fields, can generate the geometry and color of a wide range of objects without 3D supervision. Due to the scarcity of diverse, captioned 3D data, prior methods only generate objects from a handful of categories, such as ShapeNet. Instead, we guide generation with image-text models pre-trained on large datasets of captioned images from the web. Our method optimizes a Neural Radiance Field from many camera views so that rendered images score highly with a target caption according to a pre-trained CLIP model. To improve fidelity and visual quality, we introduce simple geometric priors, including sparsity-inducing transmittance regularization, scene bounds, and new MLP architectures. In experiments, Dream Fields produce realistic, multi-view consistent object geometry and color from a variety of natural language captions.

Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall et al.

Neural Radiance Fields (NeRF) have emerged as a powerful representation for the task of novel view synthesis due to their simplicity and state-of-the-art performance. Though NeRF can produce photorealistic renderings of unseen viewpoints when many input views are available, its performance drops significantly when this number is reduced. We observe that the majority of artifacts in sparse input scenarios are caused by errors in the estimated scene geometry, and by divergent behavior at the start of training. We address this by regularizing the geometry and appearance of patches rendered from unobserved viewpoints, and annealing the ray sampling space during training. We additionally use a normalizing flow model to regularize the color of unobserved viewpoints. Our model outperforms not only other methods that optimize over a single scene, but in many cases also conditional models that are extensively pre-trained on large multi-view datasets.

Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla et al.

Neural Radiance Fields (NeRF) is a technique for high quality novel view synthesis from a collection of posed input images. Like most view synthesis methods, NeRF uses tonemapped low dynamic range (LDR) as input; these images have been processed by a lossy camera pipeline that smooths detail, clips highlights, and distorts the simple noise distribution of raw sensor data. We modify NeRF to instead train directly on linear raw images, preserving the scene's full dynamic range. By rendering raw output images from the resulting NeRF, we can perform novel high dynamic range (HDR) view synthesis tasks. In addition to changing the camera viewpoint, we can manipulate focus, exposure, and tonemapping after the fact. Although a single raw image appears significantly more noisy than a postprocessed one, we show that NeRF is highly robust to the zero-mean distribution of raw noise. When optimized over many noisy raw inputs (25-200), NeRF produces a scene representation so accurate that its rendered novel views outperform dedicated single and multi-image deep raw denoisers run on the same wide baseline input images. As a result, our method, which we call RawNeRF, can reconstruct scenes from extremely noisy images captured in near-darkness.

Jonathan T. Barron, Ben Mildenhall, Dor Verbin et al.

Though neural radiance fields (NeRF) have demonstrated impressive view synthesis results on objects and small bounded regions of space, they struggle on "unbounded" scenes, where the camera may point in any direction and content may exist at any distance. In this setting, existing NeRF-like models often produce blurry or low-resolution renderings (due to the unbalanced detail and scale of nearby and distant objects), are slow to train, and may exhibit artifacts due to the inherent ambiguity of the task of reconstructing a large scene from a small set of images. We present an extension of mip-NeRF (a NeRF variant that addresses sampling and aliasing) that uses a non-linear scene parameterization, online distillation, and a novel distortion-based regularizer to overcome the challenges presented by unbounded scenes. Our model, which we dub "mip-NeRF 360" as we target scenes in which the camera rotates 360 degrees around a point, reduces mean-squared error by 57% compared to mip-NeRF, and is able to produce realistic synthesized views and detailed depth maps for highly intricate, unbounded real-world scenes.

Ayush Tewari, Justus Thies, Ben Mildenhall et al.

Synthesizing photo-realistic images and videos is at the heart of computer graphics and has been the focus of decades of research. Traditionally, synthetic images of a scene are generated using rendering algorithms such as rasterization or ray tracing, which take specifically defined representations of geometry and material properties as input. Collectively, these inputs define the actual scene and what is rendered, and are referred to as the scene representation (where a scene consists of one or more objects). Example scene representations are triangle meshes with accompanied textures (e.g., created by an artist), point clouds (e.g., from a depth sensor), volumetric grids (e.g., from a CT scan), or implicit surface functions (e.g., truncated signed distance fields). The reconstruction of such a scene representation from observations using differentiable rendering losses is known as inverse graphics or inverse rendering. Neural rendering is closely related, and combines ideas from classical computer graphics and machine learning to create algorithms for synthesizing images from real-world observations. Neural rendering is a leap forward towards the goal of synthesizing photo-realistic image and video content. In recent years, we have seen immense progress in this field through hundreds of publications that show different ways to inject learnable components into the rendering pipeline. This state-of-the-art report on advances in neural rendering focuses on methods that combine classical rendering principles with learned 3D scene representations, often now referred to as neural scene representations. A key advantage of these methods is that they are 3D-consistent by design, enabling applications such as novel viewpoint synthesis of a captured scene. In addition to methods that handle static scenes, we cover neural scene representations for modeling non-rigidly deforming objects...

Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall et al.

Neural volumetric representations such as Neural Radiance Fields (NeRF) have emerged as a compelling technique for learning to represent 3D scenes from images with the goal of rendering photorealistic images of the scene from unobserved viewpoints. However, NeRF's computational requirements are prohibitive for real-time applications: rendering views from a trained NeRF requires querying a multilayer perceptron (MLP) hundreds of times per ray. We present a method to train a NeRF, then precompute and store (i.e. "bake") it as a novel representation called a Sparse Neural Radiance Grid (SNeRG) that enables real-time rendering on commodity hardware. To achieve this, we introduce 1) a reformulation of NeRF's architecture, and 2) a sparse voxel grid representation with learned feature vectors. The resulting scene representation retains NeRF's ability to render fine geometric details and view-dependent appearance, is compact (averaging less than 90 MB per scene), and can be rendered in real-time (higher than 30 frames per second on a laptop GPU). Actual screen captures are shown in our video.

Jonathan T. Barron, Ben Mildenhall, Matthew Tancik et al.

The rendering procedure used by neural radiance fields (NeRF) samples a scene with a single ray per pixel and may therefore produce renderings that are excessively blurred or aliased when training or testing images observe scene content at different resolutions. The straightforward solution of supersampling by rendering with multiple rays per pixel is impractical for NeRF, because rendering each ray requires querying a multilayer perceptron hundreds of times. Our solution, which we call "mip-NeRF" (a la "mipmap"), extends NeRF to represent the scene at a continuously-valued scale. By efficiently rendering anti-aliased conical frustums instead of rays, mip-NeRF reduces objectionable aliasing artifacts and significantly improves NeRF's ability to represent fine details, while also being 7% faster than NeRF and half the size. Compared to NeRF, mip-NeRF reduces average error rates by 17% on the dataset presented with NeRF and by 60% on a challenging multiscale variant of that dataset that we present. Mip-NeRF is also able to match the accuracy of a brute-force supersampled NeRF on our multiscale dataset while being 22x faster.

Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang et al.

We present a method that takes as input a set of images of a scene illuminated by unconstrained known lighting, and produces as output a 3D representation that can be rendered from novel viewpoints under arbitrary lighting conditions. Our method represents the scene as a continuous volumetric function parameterized as MLPs whose inputs are a 3D location and whose outputs are the following scene properties at that input location: volume density, surface normal, material parameters, distance to the first surface intersection in any direction, and visibility of the external environment in any direction. Together, these allow us to render novel views of the object under arbitrary lighting, including indirect illumination effects. The predicted visibility and surface intersection fields are critical to our model's ability to simulate direct and indirect illumination during training, because the brute-force techniques used by prior work are intractable for lighting conditions outside of controlled setups with a single light. Our method outperforms alternative approaches for recovering relightable 3D scene representations, and performs well in complex lighting settings that have posed a significant challenge to prior work.

Matthew Tancik, Ben Mildenhall, Terrance Wang et al.

Coordinate-based neural representations have shown significant promise as an alternative to discrete, array-based representations for complex low dimensional signals. However, optimizing a coordinate-based network from randomly initialized weights for each new signal is inefficient. We propose applying standard meta-learning algorithms to learn the initial weight parameters for these fully-connected networks based on the underlying class of signals being represented (e.g., images of faces or 3D models of chairs). Despite requiring only a minor change in implementation, using these learned initial weights enables faster convergence during optimization and can serve as a strong prior over the signal class being modeled, resulting in better generalization when only partial observations of a given signal are available. We explore these benefits across a variety of tasks, including representing 2D images, reconstructing CT scans, and recovering 3D shapes and scenes from 2D image observations.

Sai Bi, Zexiang Xu, Pratul Srinivasan et al.

We present Neural Reflectance Fields, a novel deep scene representation that encodes volume density, normal and reflectance properties at any 3D point in a scene using a fully-connected neural network. We combine this representation with a physically-based differentiable ray marching framework that can render images from a neural reflectance field under any viewpoint and light. We demonstrate that neural reflectance fields can be estimated from images captured with a simple collocated camera-light setup, and accurately model the appearance of real-world scenes with complex geometry and reflectance. Once estimated, they can be used to render photo-realistic images under novel viewpoint and (non-collocated) lighting conditions and accurately reproduce challenging effects like specularities, shadows and occlusions. This allows us to perform high-quality view synthesis and relighting that is significantly better than previous methods. We also demonstrate that we can compose the estimated neural reflectance field of a real scene with traditional scene models and render them using standard Monte Carlo rendering engines. Our work thus enables a complete pipeline from high-quality and practical appearance acquisition to 3D scene composition and rendering.

Kai-En Lin, Zexiang Xu, Ben Mildenhall et al.

We propose a learning-based approach for novel view synthesis for multi-camera 360$^{\circ}$ panorama capture rigs. Previous work constructs RGBD panoramas from such data, allowing for view synthesis with small amounts of translation, but cannot handle the disocclusions and view-dependent effects that are caused by large translations. To address this issue, we present a novel scene representation - Multi Depth Panorama (MDP) - that consists of multiple RGBD$α$ panoramas that represent both scene geometry and appearance. We demonstrate a deep neural network-based method to reconstruct MDPs from multi-camera 360$^{\circ}$ images. MDPs are more compact than previous 3D scene representations and enable high-quality, efficient new view rendering. We demonstrate this via experiments on both synthetic and real data and comparisons with previous state-of-the-art methods spanning both learning-based approaches and classical RGBD-based methods.

Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall et al.

We show that passing input points through a simple Fourier feature mapping enables a multilayer perceptron (MLP) to learn high-frequency functions in low-dimensional problem domains. These results shed light on recent advances in computer vision and graphics that achieve state-of-the-art results by using MLPs to represent complex 3D objects and scenes. Using tools from the neural tangent kernel (NTK) literature, we show that a standard MLP fails to learn high frequencies both in theory and in practice. To overcome this spectral bias, we use a Fourier feature mapping to transform the effective NTK into a stationary kernel with a tunable bandwidth. We suggest an approach for selecting problem-specific Fourier features that greatly improves the performance of MLPs for low-dimensional regression tasks relevant to the computer vision and graphics communities.

Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik et al.

We present a method that achieves state-of-the-art results for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function using a sparse set of input views. Our algorithm represents a scene using a fully-connected (non-convolutional) deep network, whose input is a single continuous 5D coordinate (spatial location $(x,y,z)$ and viewing direction $(θ, φ)$) and whose output is the volume density and view-dependent emitted radiance at that spatial location. We synthesize views by querying 5D coordinates along camera rays and use classic volume rendering techniques to project the output colors and densities into an image. Because volume rendering is naturally differentiable, the only input required to optimize our representation is a set of images with known camera poses. We describe how to effectively optimize neural radiance fields to render photorealistic novel views of scenes with complicated geometry and appearance, and demonstrate results that outperform prior work on neural rendering and view synthesis. View synthesis results are best viewed as videos, so we urge readers to view our supplementary video for convincing comparisons.

Pratul P. Srinivasan, Ben Mildenhall, Matthew Tancik et al.

We present a deep learning solution for estimating the incident illumination at any 3D location within a scene from an input narrow-baseline stereo image pair. Previous approaches for predicting global illumination from images either predict just a single illumination for the entire scene, or separately estimate the illumination at each 3D location without enforcing that the predictions are consistent with the same 3D scene. Instead, we propose a deep learning model that estimates a 3D volumetric RGBA model of a scene, including content outside the observed field of view, and then uses standard volume rendering to estimate the incident illumination at any 3D location within that volume. Our model is trained without any ground truth 3D data and only requires a held-out perspective view near the input stereo pair and a spherical panorama taken within each scene as supervision, as opposed to prior methods for spatially-varying lighting estimation, which require ground truth scene geometry for training. We demonstrate that our method can predict consistent spatially-varying lighting that is convincing enough to plausibly relight and insert highly specular virtual objects into real images.

Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon et al.

We present a practical and robust deep learning solution for capturing and rendering novel views of complex real world scenes for virtual exploration. Previous approaches either require intractably dense view sampling or provide little to no guidance for how users should sample views of a scene to reliably render high-quality novel views. Instead, we propose an algorithm for view synthesis from an irregular grid of sampled views that first expands each sampled view into a local light field via a multiplane image (MPI) scene representation, then renders novel views by blending adjacent local light fields. We extend traditional plenoptic sampling theory to derive a bound that specifies precisely how densely users should sample views of a given scene when using our algorithm. In practice, we apply this bound to capture and render views of real world scenes that achieve the perceptual quality of Nyquist rate view sampling while using up to 4000x fewer views. We demonstrate our approach's practicality with an augmented reality smartphone app that guides users to capture input images of a scene and viewers that enable realtime virtual exploration on desktop and mobile platforms.

Matthew Tancik, Ben Mildenhall, Ren Ng

Printed and digitally displayed photos have the ability to hide imperceptible digital data that can be accessed through internet-connected imaging systems. Another way to think about this is physical photographs that have unique QR codes invisibly embedded within them. This paper presents an architecture, algorithms, and a prototype implementation addressing this vision. Our key technical contribution is StegaStamp, a learned steganographic algorithm to enable robust encoding and decoding of arbitrary hyperlink bitstrings into photos in a manner that approaches perceptual invisibility. StegaStamp comprises a deep neural network that learns an encoding/decoding algorithm robust to image perturbations approximating the space of distortions resulting from real printing and photography. We demonstrates real-time decoding of hyperlinks in photos from in-the-wild videos that contain variation in lighting, shadows, perspective, occlusion and viewing distance. Our prototype system robustly retrieves 56 bit hyperlinks after error correction - sufficient to embed a unique code within every photo on the internet.

Tim Brooks, Ben Mildenhall, Tianfan Xue et al.

Machine learning techniques work best when the data used for training resembles the data used for evaluation. This holds true for learned single-image denoising algorithms, which are applied to real raw camera sensor readings but, due to practical constraints, are often trained on synthetic image data. Though it is understood that generalizing from synthetic to real data requires careful consideration of the noise properties of image sensors, the other aspects of a camera's image processing pipeline (gain, color correction, tone mapping, etc) are often overlooked, despite their significant effect on how raw measurements are transformed into finished images. To address this, we present a technique to "unprocess" images by inverting each step of an image processing pipeline, thereby allowing us to synthesize realistic raw sensor measurements from commonly available internet photos. We additionally model the relevant components of an image processing pipeline when evaluating our loss function, which allows training to be aware of all relevant photometric processing that will occur after denoising. By processing and unprocessing model outputs and training data in this way, we are able to train a simple convolutional neural network that has 14%-38% lower error rates and is 9x-18x faster than the previous state of the art on the Darmstadt Noise Dataset, and generalizes to sensors outside of that dataset as well.

Ben Mildenhall, Jonathan T. Barron, Jiawen Chen et al.

We present a technique for jointly denoising bursts of images taken from a handheld camera. In particular, we propose a convolutional neural network architecture for predicting spatially varying kernels that can both align and denoise frames, a synthetic data generation approach based on a realistic noise formation model, and an optimization guided by an annealed loss function to avoid undesirable local minima. Our model matches or outperforms the state-of-the-art across a wide range of noise levels on both real and synthetic data.

Nick Antipa, Grace Kuo, Reinhard Heckel et al.

We demonstrate a compact and easy-to-build computational camera for single-shot 3D imaging. Our lensless system consists solely of a diffuser placed in front of a standard image sensor. Every point within the volumetric field-of-view projects a unique pseudorandom pattern of caustics on the sensor. By using a physical approximation and simple calibration scheme, we solve the large-scale inverse problem in a computationally efficient way. The caustic patterns enable compressed sensing, which exploits sparsity in the sample to solve for more 3D voxels than pixels on the 2D sensor. Our 3D voxel grid is chosen to match the experimentally measured two-point optical resolution across the field-of-view, resulting in 100 million voxels being reconstructed from a single 1.3 megapixel image. However, the effective resolution varies significantly with scene content. Because this effect is common to a wide range of computational cameras, we provide new theory for analyzing resolution in such systems.