We address the problem of generating videos from unposed internet photos. A handful of input images serve as keyframes, and our model interpolates between them to simulate a path moving between the cameras. Given random images, a model’s ability to capture underlying geometry, recognize scene identity, and relate frames in terms of camera position and orientation reflects a fundamental understanding of 3D structure and scene layout. However, existing video models such as Luma Dream Machine fail at this task. We design a self-supervised method that takes advantage of the consistency of videos and variability of multiview internet photos to train a scalable, 3D-aware video model without any 3D annotations such as camera parameters. We validate that our method outperforms all baselines in terms of geometric and appearance consistency. We also show our model benefits applications that enable camera control, such as 3D Gaussian Splatting. Our results suggest that we can scale up scene-level 3D learning using only 2D data such as videos and multiview internet photos.

We present Turbo3D, an ultra-fast text-to-3D system capable of generating high-quality Gaussian splatting assets in under one second. Turbo3D employs a rapid 4-step, 4-view diffusion generator and an efficient feed-forward Gaussian reconstructor, both operating in latent space, as shown in Fig. 1. The 4-step, 4-view generator is a student model distilled through a novel Dual-Teacher approach, which encourages the student to learn view consistency from a multiview teacher and photo-realism from a single-view teacher. By shifting the Gaussian reconstructor’s inputs from pixel space to latent space, we eliminate the extra image decoding time and halve the transformer sequence length for maximum efficiency. Our method demonstrates superior 3D generation results compared to previous baselines, while operating in a fraction of their runtime.

We propose scaling up 3D scene reconstruction by training with synthesized data. At the core of our work is MegaSynth, a procedurally generated 3D dataset comprising 700K scenes—over 50 times larger than the prior real dataset DL3DV—dramatically scaling the training data. To enable scalable data generation, our key idea is eliminating semantic information, removing the need to model complex semantic priors such as object affordances and scene composition. Instead, we model scenes with basic spatial structures and geometry primitives, offering scalability. Besides, we control data complexity to facilitate training while loosely aligning it with real-world data distribution to benefit real-world generalization. We explore training LRMs with both MegaSynth and available real data. Experiment results show that joint training or pre-training with MegaSynth improves reconstruction quality by 1.2 to 1.8 dB PSNR across diverse image domains. Moreover, models trained solely on MegaSynth perform comparably to those trained on real data, underscoring the low-level nature of 3D reconstruction. Additionally, we provide an in-depth analysis of MegaSynth’s properties for enhancing model capability, training stability, and generalization, as well as application to other tasks.

We present Buffer Anytime, a framework for estimation of depth and normal maps (which we call geometric buffers) from video that eliminates the need for paired video–depth and video–normal training data. Instead of relying on large-scale annotated video datasets, we demonstrate high-quality video buffer estimation by leveraging single-image priors with temporal consistency constraints. Our zero-shot training strategy combines state-of-the-art image estimation models based on optical flow smoothness through a hybrid loss function, implemented via a lightweight temporal attention architecture. Applied to leading image models like Depth Anything V2 and Marigold-E2E-FT, our approach significantly improves temporal consistency while maintaining accuracy. Experiments show that our method not only outperforms image-based approaches but also achieves results comparable to state-of-the-art video models trained on large-scale paired video datasets, despite using no such paired video data.

We propose Long-LRM, a feed-forward 3D Gaussian reconstruction model for instant, high-resolution, 360° wide-coverage, scene-level reconstruction. Specifically, it takes in 32 input images at a resolution of 960x540 and produces the Gaussian reconstruction in just 1 second on a single A100 GPU. To handle the long sequence of 250K tokens brought by the large input size, Long-LRM features a mixture of the recent Mamba2 blocks and the classical transformer blocks, enhanced by a light-weight token merging module and Gaussian pruning steps that balance between quality and efficiency. We evaluate Long-LRM on the large-scale DL3DV benchmark and Tanks&Temples, demonstrating reconstruction quality comparable to the optimization-based methods while achieving an 800x speedup w.r.t. the optimization-based approaches and an input size at least 60x larger than the previous feed-forward approaches. We conduct extensive ablation studies on our model design choices for both rendering quality and computation efficiency. We also explore Long-LRM's compatibility with other Gaussian variants such as 2D GS, which enhances Long-LRM's ability in geometry reconstruction. Project page: https://arthurhero.github.io/projects/llrm

Recent probabilistic methods for 3D triangular meshes capture diverse shapes by differentiable mesh connectivity, but face high computational costs with increased shape details. We introduce a new differentiable mesh processing method in 2D and 3D that addresses this challenge and efficiently handles meshes with intricate structures. Additionally, we present an algorithm that adapts the mesh resolution to local geometry in 2D for efficient representation. We demonstrate the effectiveness of our approach on 2D point cloud and 3D multi-view reconstruction tasks. Visit our project page (https://sonsang.github.io/dmesh2-project) for source code and supplementary material.

We propose the Large View Synthesis Model (LVSM), a novel transformer-based approach for scalable and generalizable novel view synthesis from sparse-view inputs. We introduce two architectures: (1) an encoder-decoder LVSM, which encodes input image tokens into a fixed number of 1D latent tokens, functioning as a fully learned scene representation, and decodes novel-view images from them; and (2) a decoder-only LVSM, which directly maps input images to novel-view outputs, completely eliminating intermediate scene representations. Both models bypass the 3D inductive biases used in previous methods---from 3D representations (e.g., NeRF, 3DGS) to network designs (e.g., epipolar projections, plane sweeps)---addressing novel view synthesis with a fully data-driven approach. While the encoder-decoder model offers faster inference due to its independent latent representation, the decoder-only LVSM achieves superior quality, scalability, and zero-shot generalization, outperforming previous state-of-the-art methods by 1.5 to 3.5 dB PSNR. Comprehensive evaluations across multiple datasets demonstrate that both LVSM variants achieve state-of-the-art novel view synthesis quality, delivering superior performance even with reduced computational resources (1-2 GPUs). Please see our anonymous website for more details: https://haian-jin.github.io/projects/LVSM/

We propose RelitLRM, a Large Reconstruction Model (LRM) for generating high-quality Gaussian splatting representations of 3D objects under novel illuminations from sparse (4-8) posed images captured under unknown static lighting. Unlike prior inverse rendering methods requiring dense captures and slow optimization, often causing artifacts like incorrect highlights or shadow baking, RelitLRM adopts a feed-forward transformer-based model with a novel combination of a geometry reconstructor and a relightable appearance generator based on diffusion. The model is trained end-to-end on synthetic multi-view renderings of objects under varying known illuminations. This architecture design enables to effectively decompose geometry and appearance, resolve the ambiguity between material and lighting, and capture the multi-modal distribution of shadows and specularity in the relit appearance. We show our sparse-view feed-forward RelitLRM offers competitive relighting results to state-of-the-art dense-view optimization-based baselines while being significantly faster. Our project page is available at: https://relit-lrm.github.io/.

Diffusion models approximate the denoising distribution as a Gaussian and predict its mean, whereas flow matching models reparameterize the Gaussian mean as flow velocity. However, they underperform in few-step sampling due to discretization error and tend to produce over-saturated colors under classifier-free guidance (CFG). To address these limitations, we propose a novel Gaussian mixture flow matching (GMFlow) model: instead of predicting the mean, GMFlow predicts dynamic Gaussian mixture (GM) parameters to capture a multi-modal flow velocity distribution, which can be learned with a KL divergence loss. We demonstrate that GMFlow generalizes previous diffusion and flow matching models where a single Gaussian is learned with an $L_2$ denoising loss. For inference, we derive GM-SDE/ODE solvers that leverage analytic denoising distributions and velocity fields for precise few-step sampling. Furthermore, we introduce a novel probabilistic guidance scheme that mitigates the over-saturation issues of CFG and improves image generation quality. Extensive experiments demonstrate that GMFlow consistently outperforms flow matching baselines in generation quality, achieving a Precision of 0.942 with only 6 sampling steps on ImageNet 256$\times$256.

Can we scale 4D pretraining to learn general space-time representations that reconstruct an object from a few views at some times to any view at any time? We provide an affirmative answer with 4D-LRM, the first large-scale 4D reconstruction model that takes input from unconstrained views and timestamps and renders arbitrary novel view-time combinations. Unlike prior 4D approaches, e.g., optimization-based, geometry-based, or generative, that struggle with efficiency, generalization, or faithfulness, 4D-LRM learns a unified space-time representation and directly predicts per-pixel 4D Gaussian primitives from posed image tokens across time, enabling fast, high-quality rendering at, in principle, infinite frame rate. Our results demonstrate that scaling spatiotemporal pretraining enables accurate and efficient 4D reconstruction. We show that 4D-LRM generalizes to novel objects, interpolates across time, and handles diverse camera setups. It reconstructs 24-frame sequences in one forward pass with less than 1.5 seconds on a single A100 GPU.

While existing volumetric rendering approaches provide photorealistic results, extracting high-quality meshes from optimized neural field representations is challenging. Conversely, existing differentiable rasterization-based methods are typically sensitive to initialization and suffer from poor mesh rendering quality. In this paper, we introduce Neu-Manifold, a novel method for reconstructing watertight manifold meshes with high-quality textures from multi-view input images. NeuManifold overcomes the limitations of existing approaches by first learning a neural volumetric field and then refining it through differentiable mesh extraction and surface rendering. To eliminate artifacts and preserve mesh properties during iso-surface extraction, we introduce a novel differentiable marching cubes method. Instead of traditional textures, we use neural textures to enhance rendering quality. To integrate with modern graphics rendering pipelines, we also provide customized GLSL shader support for neural textures. Extensive experiments demonstrate that NeuManifold outperforms existing mesh-based reconstruction methods in both mesh quality and rendering metrics, achieving comparable or superior rendering quality to prior volume-rendering-based methods. The generated results enable real-time, high-quality rendering and seamlessly support numerous graphics pipelines and applications requiring high-quality meshes, such as 3D printing and physical simulation. https://sarahweiii.github.io/neumanifold/.

We present RigAnything , a novel autoregressive transformer-based model, which makes 3D assets rig-ready by probabilistically generating joints and skeleton topologies and assigning skinning weights in a template-free manner. Unlike most existing auto-rigging methods, which rely on predefined skeleton templates and are limited to specific categories like humanoid, RigAnything approaches the rigging problem in an autoregressive manner, iteratively predicting the next joint based on the global input shape and the previous prediction. While autoregressive models are typically used to generate sequential data, RigAnything extends its application to effectively learn and represent skeletons, which are inherently tree structures. To achieve this, we organize the joints in a breadth-first search (BFS) order, enabling the skeleton to be defined as a sequence of 3D locations and the parent index. Furthermore, our model improves the accuracy of position prediction by leveraging diffusion modeling, ensuring precise and consistent placement of joints within the hierarchy. This formulation allows the autoregressive model to efficiently capture both spatial and hierarchical relationships within the skeleton. Trained end-to-end on both RigNet and Objaverse datasets, RigAnything demonstrates state-of-the-art performance across diverse object types, including humanoids, quadrupeds, marine creatures, insects, and many more, surpassing prior methods in quality, robustness, generalizability, and efficiency. It achieves significantly faster performance than existing auto-rigging methods, completing rigging in under a few seconds per shape.

Multi-view mesh reconstruction has long been a challenging problem in graphics and computer vision. In contrast to recent volumetric rendering methods that generate meshes through post-processing, we propose an end-to-end mesh optimization approach called IMLS-Splatting. Our method leverages the sparsity and flexibility of point clouds to efficiently represent the underlying surface. To achieve this, we introduce a splatting-based differentiable Implicit Moving-Least Squares (IMLS) algorithm that enables the fast conversion of point clouds into SDFs and texture fields, optimizing both mesh reconstruction and rasterization. Additionally, the IMLS representation ensures that the reconstructed SDF and mesh maintain continuity and smoothness without the need for extra regularization. With this efficient pipeline, our method enables the reconstruction of highly detailed meshes in approximately 11 minutes, supporting high-quality rendering and achieving state-of-the-art reconstruction performance. Our code is available at https://github.com/SilenKZYoung/IMLS-Splatting.

Recent advances in Neural Radiance Fields (NeRFs) have made it possible to reconstruct and reanimate dynamic portrait scenes with control over head-pose, facial expressions and viewing direction. However, training such models assumes photometric consistency over the deformed region e.g. the face must be evenly lit as it deforms with changing head-pose and facial expression. Such photometric consistency across frames of a video is hard to maintain, even in studio environments, thus making the created reanimatable neural portraits prone to artefacts during reanimation. In this work, we propose CoDyNeRF, a system that enables the creation of fully controllable 3D portraits in real-world capture conditions. CoDyNeRF learns to approximate illumination dependent effects via a dynamic appearance model in the canonical space that is conditioned on predicted surface normals and the facial expressions and head-pose deformations. The surface normals prediction is guided using 3DMM normals that act as a coarse prior for the normals of the human head, where direct prediction of normals is hard due to rigid and non-rigid deformations induced by head-pose and facial expression changes. Using only a smartphone-captured short video of a subject for training, we demonstrate the effectiveness of our method on free view synthesis of a portrait scene with explicit head pose and expression controls, and realistic lighting effects.

Novel-view synthesis of specular objects like shiny metals or glossy paints remains a significant challenge. Not only the glossy appearance but also global illumination effects, including reflections of other objects in the environment, are critical components to faithfully reproduce a scene. In this paper, we present Neural Directional Encoding (NDE), a view-dependent appearance encoding of neural radiance fields (NeRF) for rendering specular objects. NDE transfers the concept of feature-grid-based spatial encoding to the angular domain, significantly improving the ability to model high-frequency angular signals. In contrast to previous methods that use encoding functions with only angular input, we additionally cone-trace spatial features to obtain a spatially varying directional encoding, which addresses the challenging interreflection effects Extensive experiments on both synthetic and real datasets show that a NeRF model with NDE (1) outperforms the state of the art on view synthesis of specular objects, and (2) works with small networks to allow fast (real-time) inference. The source code is available at: https://github.com/lwwu2/nde

We propose a Pose-Free Large Reconstruction Model (PF-LRM) for reconstructing a 3D object from a few unposed images even with little visual overlap, while simultaneously estimating the relative camera poses in ~1.3 seconds on a single A100 GPU. PF-LRM is a highly scalable method utilizing self-attention blocks to exchange information between 3D object tokens and 2D image tokens; we predict a coarse point cloud for each view, and then use a differentiable Perspective-n-Point (PnP) solver to obtain camera poses. When trained on a huge amount of multi-view posed data of ~1M objects, PF-LRM shows strong cross-dataset generalization ability, and outperforms baseline methods by a large margin in terms of pose prediction accuracy and 3D reconstruction quality on various unseen evaluation datasets. We also demonstrate our model's applicability in downstream text/image-to-3D task with fast feed-forward inference. Our project website is at: https://totoro97.github.io/pf-lrm.

Text-to-3D with diffusion models has achieved remarkable progress in recent years. However, existing methods either rely on score distillation-based optimization which suffer from slow inference, low diversity and Janus problems, or are feed-forward methods that generate low-quality results due to the scarcity of 3D training data. In this paper, we propose Instant3D, a novel method that generates high-quality and diverse 3D assets from text prompts in a feed-forward manner. We adopt a two-stage paradigm, which first generates a sparse set of four structured and consistent views from text in one shot with a fine-tuned 2D text-to-image diffusion model, and then directly regresses the NeRF from the generated images with a novel transformer-based sparse-view reconstructor. Through extensive experiments, we demonstrate that our method can generate diverse 3D assets of high visual quality within 20 seconds, which is two orders of magnitude faster than previous optimization-based methods that can take 1 to 10 hours. Our project webpage is: https://jiahao.ai/instant3d/.

We propose DMV3D, a novel 3D generation approach that uses a transformer-based 3D large reconstruction model to denoise multi-view diffusion. Our reconstruction model incorporates a triplane NeRF representation and, functioning as a denoiser, can denoise noisy multi-view images via 3D NeRF reconstruction and rendering, achieving single-stage 3D generation in the 2D diffusion denoising process. We train DMV3D on large-scale multi-view image datasets of extremely diverse objects using only image reconstruction losses, without accessing 3D assets. We demonstrate state-of-the-art results for the single-image reconstruction problem where probabilistic modeling of unseen object parts is required for generating diverse reconstructions with sharp textures. We also show high-quality text-to-3D generation results outperforming previous 3D diffusion models. Our project website is at: https://dmv3d.github.io/.

We present MovingParts, a NeRF-based method for dynamic scene reconstruction and part discovery. We consider motion as an important cue for identifying parts, that all particles on the same part share the common motion pattern. From the perspective of fluid simulation, existing deformation-based methods for dynamic NeRF can be seen as parameterizing the scene motion under the Eulerian view, i.e., focusing on specific locations in space through which the fluid flows as time passes. However, it is intractable to extract the motion of constituting objects or parts using the Eulerian view representation. In this work, we introduce the dual Lagrangian view and enforce representations under the Eulerian/Lagrangian views to be cycle-consistent. Under the Lagrangian view, we parameterize the scene motion by tracking the trajectory of particles on objects. The Lagrangian view makes it convenient to discover parts by factorizing the scene motion as a composition of part-level rigid motions. Experimentally, our method can achieve fast and high-quality dynamic scene reconstruction from even a single moving camera, and the induced part-based representation allows direct applications of part tracking, animation, 3D scene editing, etc.

We present a differentiable representation, DMesh, for general 3D triangular meshes. DMesh considers both the geometry and connectivity information of a mesh. In our design, we first get a set of convex tetrahedra that compactly tessellates the domain based on Weighted Delaunay Triangulation (WDT), and select triangular faces on the tetrahedra to define the final mesh. We formulate probability of faces to exist on the actual surface in a differentiable manner based on the WDT. This enables DMesh to represent meshes of various topology in a differentiable way, and allows us to reconstruct the mesh under various observations, such as point clouds and multi-view images using gradient-based optimization. We publicize the source code and supplementary material at our project page (https://sonsang.github.io/dmesh-project).

Single-image relighting is a challenging task that involves reasoning about the complex interplay between geometry, materials, and lighting. Many prior methods either support only specific categories of images, such as portraits, or require special capture conditions, like using a flashlight. Alternatively, some methods explicitly decompose a scene into intrinsic components, such as normals and BRDFs, which can be inaccurate or under-expressive. In this work, we propose a novel end-to-end 2D relighting diffusion model, called Neural Gaffer, that takes a single image of any object and can synthesize an accurate, high-quality relit image under any novel environmental lighting condition, simply by conditioning an image generator on a target environment map, without an explicit scene decomposition. Our method builds on a pre-trained diffusion model, and fine-tunes it on a synthetic relighting dataset, revealing and harnessing the inherent understanding of lighting present in the diffusion model. We evaluate our model on both synthetic and in-the-wild Internet imagery and demonstrate its advantages in terms of generalization and accuracy. Moreover, by combining with other generative methods, our model enables many downstream 2D tasks, such as text-based relighting and object insertion. Our model can also operate as a strong relighting prior for 3D tasks, such as relighting a radiance field.

We present LRM-Zero, a Large Reconstruction Model (LRM) trained entirely on synthesized 3D data, achieving high-quality sparse-view 3D reconstruction. The core of LRM-Zero is our procedural 3D dataset, Zeroverse, which is automatically synthesized from simple primitive shapes with random texturing and augmentations (e.g., height fields, boolean differences, and wireframes). Unlike previous 3D datasets (e.g., Objaverse) which are often captured or crafted by humans to approximate real 3D data, Zeroverse completely ignores realistic global semantics but is rich in complex geometric and texture details that are locally similar to or even more intricate than real objects. We demonstrate that our LRM-Zero, trained with our fully synthesized Zeroverse, can achieve high visual quality in the reconstruction of real-world objects, competitive with models trained on Objaverse. We also analyze several critical design choices of Zeroverse that contribute to LRM-Zero's capability and training stability. Our work demonstrates that 3D reconstruction, one of the core tasks in 3D vision, can potentially be addressed without the semantics of real-world objects. The Zeroverse's procedural synthesis code and interactive visualization are available at: https://desaixie.github.io/lrm-zero/.

Diffusion models currently achieve state-of-the-art performance for both conditional and unconditional image generation. However, so far, image diffusion models do not support tasks required for 3D understanding, such as view-consistent 3D generation or single-view object reconstruction. In this paper, we present RenderDiffusion, the first diffusion model for 3D generation and inference, trained using only monocular 2D supervision. Central to our method is a novel image denoising architecture that generates and renders an intermediate three-dimensional representation of a scene in each denoising step. This enforces a strong inductive structure within the diffusion process, providing a 3D consistent representation while only requiring 2D supervision. The resulting 3D representation can be rendered from any view. We evaluate RenderDiffusion on FFHQ, AFHQ, ShapeNet and CLEVR datasets, showing competitive performance for generation of 3D scenes and inference of 3D scenes from 2D images. Additionally, our diffusion-based approach allows us to use 2D inpainting to edit 3D scenes.

We propose TensoIR, a novel inverse rendering approach based on tensor factorization and neural fields. Unlike previous works that use purely MLP-based neural fields, thus suffering from low capacity and high computation costs, we extend TensoRF, a state-of-the-art approach for radiance field modeling, to estimate scene geometry, surface reflectance, and environment illumination from multi-view images captured under unknown lighting conditions. Our approach jointly achieves radiance field reconstruction and physically-based model estimation, leading to photo-realistic novel view synthesis and relighting results. Benefiting from the efficiency and extensibility of the TensoRF-based representation, our method can accurately model secondary shading effects (like shadows and indirect lighting) and generally support input images captured under single or multiple unknown lighting conditions. The low-rank tensor representation allows us to not only achieve fast and compact reconstruction but also better exploit shared information under an arbitrary number of capturing lighting conditions. We demonstrate the superiority of our method to baseline methods qualitatively and quantitatively on various challenging synthetic and real-world scenes.

We propose Strivec, a novel neural representation that models a 3D scene as a radiance field with sparsely distributed and compactly factorized local tensor feature grids. Our approach leverages tensor decomposition, following the recent work TensoRF [6], to model the tensor grids. In contrast to TensoRF which uses a global tensor and focuses on their vector-matrix decomposition, we propose to utilize a cloud of local tensors and apply the classic CANDE-COMP/PARAFAC (CP) decomposition [4] to factorize each tensor into triple vectors that express local feature distributions along spatial axes and compactly encode a local neural field. We also apply multi-scale tensor grids to discover the geometry and appearance commonalities and exploit spatial coherence with the tri-vector factorization at multiple local scales. The final radiance field properties are regressed by aggregating neural features from multiple local tensors across all scales. Our tri-vector tensors are sparsely distributed around the actual scene surface, discovered by a fast coarse reconstruction, leveraging the sparsity of a 3D scene. We demonstrate that our model can achieve better rendering quality while using significantly fewer parameters than previous methods, including TensoRF and Instant-NGP [23].

We introduce OpenIllumination, a real-world dataset containing over 108K images of 64 objects with diverse materials, captured under 72 camera views and a large number of different illuminations. For each image in the dataset, we provide accurate camera parameters, illumination ground truth, and foreground segmentation masks. Our dataset enables the quantitative evaluation of most inverse rendering and material decomposition methods for real objects. We examine several state-of-the-art inverse rendering methods on our dataset and compare their performances. The dataset and code can be found on the project page: https://oppo-us-research.github.io/OpenIllumination.

Single image 3D reconstruction is an important but challenging task that requires extensive knowledge of our natural world. Many existing methods solve this problem by optimizing a neural radiance field under the guidance of 2D diffusion models but suffer from lengthy optimization time, 3D inconsistency results, and poor geometry. In this work, we propose a novel method that takes a single image of any object as input and generates a full 360-degree 3D textured mesh in a single feed-forward pass. Given a single image, we first use a view-conditioned 2D diffusion model, Zero123, to generate multi-view images for the input view, and then aim to lift them up to 3D space. Since traditional reconstruction methods struggle with inconsistent multi-view predictions, we build our 3D reconstruction module upon an SDF-based generalizable neural surface reconstruction method and propose several critical training strategies to enable the reconstruction of 360-degree meshes. Without costly optimizations, our method reconstructs 3D shapes in significantly less time than existing methods. Moreover, our method favors better geometry, generates more 3D consistent results, and adheres more closely to the input image. We evaluate our approach on both synthetic data and in-the-wild images and demonstrate its superiority in terms of both mesh quality and runtime. In addition, our approach can seamlessly support the text-to-3D task by integrating with off-the-shelf text-to-image diffusion models.

The volume rendering step used in Neural Radiance Fields (NeRFs) produces highly photorealistic results, but is inherently slow because it evaluates an MLP at a large number of sample points per ray. Previous work has addressed this by either proposing neural scene representations that are faster to evaluate or by pre-computing (and approximating) scene properties to reduce render times. In this work, we propose MCNeRF, a general Monte Carlo-based rendering algorithm that can speed up any NeRF representation. We show that the NeRF volume rendering integral can be efficiently computed via Monte Carlo integration using an importance sampling scheme based on ray density distributions. This allows us to use a small number of MLP evaluations to estimate pixel radiance. These noisy Monte Carlo estimates can be further denoised using an inexpensive image-space denoiser trained per-scene. We demonstrate that MCNeRF can be used to speed up NeRF representations like TensoRF by 7 × while closely matching their visual quality and without making the scene approximations that real-time NeRF rendering methods usually make.

Abstract Precomputed Radiance Transfer (PRT) remains an attractive solution for real‐time rendering of complex light transport effects such as glossy global illumination. After precomputation, we can relight the scene with new environment maps while changing viewpoint in real‐time. However, practical PRT methods are usually limited to low‐frequency spherical harmonic lighting. All‐frequency techniques using wavelets are promising but have so far had little practical impact. The curse of dimensionality and much higher data requirements have typically limited them to relighting with fixed view or only direct lighting with triple product integrals. In this paper, we demonstrate a hybrid neural‐wavelet PRT solution to high‐frequency indirect illumination, including glossy reflection, for relighting with changing view. Specifically, we seek to represent the light transport function in the Haar wavelet basis. For global illumination, we learn the wavelet transport using a small multi‐layer perceptron (MLP) applied to a feature field as a function of spatial location and wavelet index, with reflected direction and material parameters being other MLP inputs. We optimize/learn the feature field (compactly represented by a tensor decomposition) and MLP parameters from multiple images of the scene under different lighting and viewing conditions. We demonstrate real‐time (512 x 512 at 24 FPS, 800 x 600 at 13 FPS) precomputed rendering of challenging scenes involving view‐dependent reflections and even caustics.

We present Dictionary Fields, a novel neural representation which decomposes a signal into a product of factors, each represented by a classical or neural field representation, operating on transformed input coordinates. More specifically, we factorize a signal into a coefficient field and a basis field, and exploit periodic coordinate transformations to apply the same basis functions across multiple locations and scales. Our experiments show that Dictionary Fields lead to improvements in approximation quality, compactness, and training time when compared to previous fast reconstruction methods. Experimentally, our representation achieves better image approximation quality on 2D image regression tasks, higher geometric quality when reconstructing 3D signed distance fields, and higher compactness for radiance field reconstruction tasks. Furthermore, Dictionary Fields enable generalization to unseen images/3D scenes by sharing bases across signals during training which greatly benefits use cases such as image regression from partial observations and few-shot radiance field reconstruction.

Volumetric neural rendering methods, such as neural radiance fields (NeRFs), have enabled photo-realistic novel view synthesis. However, in their standard form, NeRFs do not support the editing of objects, such as a human head, within a scene. In this work, we propose RigNeRF, a system that goes beyond just novel view synthesis and enables full control of head pose and facial expressions learned from a single portrait video. We model changes in head pose and facial expressions using a deformation field that is guided by a 3D morphable face model (3DMM). The 3DMM effectively acts as a prior for RigNeRF that learns to predict only residuals to the 3DMM deformations and allows us to render novel (rigid) poses and (non-rigid) expressions that were not present in the input sequence. Using only a smartphone-captured short video of a subject for training, we demonstrate the effectiveness of our method on free view synthesis of a portrait scene with explicit head pose and expression controls.

Volumetric neural rendering methods like NeRF [34] generate high-quality view synthesis results but are optimized per-scene leading to prohibitive reconstruction time. On the other hand, deep multi-view stereo methods can quickly reconstruct scene geometry via direct network inference. Point-NeRF combines the advantages of these two approaches by using neural 3D point clouds, with associated neural features, to model a radiance field. Point-NeRF can be rendered efficiently by aggregating neural point features near scene surfaces, in a ray marching-based rendering pipeline. Moreover, Point-NeRF can be initialized via direct inference of a pre-trained deep network to produce a neural point cloud; this point cloud can be finetuned to surpass the visual quality of NeRF with 30× faster training time. Point-NeRF can be combined with other 3D re-construction methods and handles the errors and outliers in such methods via a novel pruning and growing mechanism.

Most indoor 3D scene reconstruction methods focus on recovering 3D geometry and scene layout. In this work, we go beyond this to propose PhotoScene11Code: https://github.com/ViLab-UCSD/PhotoScene, a framework that takes input image(s) of a scene along with approximately aligned CAD geometry (either reconstructed automatically or manually specified) and builds a photorealistic digital twin with high-quality materials and similar lighting. We model scene materials using procedural material graphs; such graphs represent photorealistic and resolution-independent materials. We optimize the parameters of these graphs and their texture scale and rotation, as well as the scene lighting to best match the input image via a differentiable rendering layer. We evaluate our technique on objects and layout reconstructions from ScanNet, SUN RGB-D and stock photographs, and demonstrate that our method reconstructs high-quality, fully relightable 3D scenes that can be re-rendered under arbitrary viewpoints, zooms and lighting.

While NeRF [28] has shown great success for neural reconstruction and rendering, its limited MLP capacity and long per-scene optimization times make it challenging to model large-scale indoor scenes. In contrast, classical 3D reconstruction methods can handle large-scale scenes but do not produce realistic renderings. We propose NeRFusion, a method that combines the advantages of NeRF and TSDF-based fusion techniques to achieve efficient large-scale reconstruction and photo-realistic rendering. We process the input image sequence to predict per-frame local radiance fields via direct network inference. These are then fused using a novel recurrent neural network that incrementally reconstructs a global, sparse scene representation in real-time at 22 fps. This global volume can be further fine-tuned to boost rendering quality. We demonstrate that NeR-Fusionachieves state-of-the-art quality on both large-scale indoor and small-scale object scenes, with substantially faster reconstruction than NeRF and other recent methods.11https://jetd1.github.io/NeRFusion-Web/

We present a method to automatically compute correct gradients with respect to geometric scene parameters in neural SDF renderers. Recent physically-based differentiable rendering techniques for meshes have used edge-sampling to handle discontinuities, particularly at object silhouettes, but SDFs do not have a simple parametric form amenable to sampling. Instead, our approach builds on area-sampling techniques and develops a continuous warping function for SDFs to account for these discontinuities. Our method leverages the distance to surface encoded in an SDF and uses quadrature on sphere tracer points to compute this warping function. We further show that this can be done by subsampling the points to make the method tractable for neural SDFs. Our differentiable renderer can be used to optimize neural shapes from multi-view images and produces comparable 3D reconstructions to recent SDF-based inverse rendering methods, without the need for 2D segmentation masks to guide the geometry optimization and no volumetric approximations to the geometry.

Although Monte Carlo path tracing is a simple and effective algorithm to synthesize photo-realistic images, it is often very slow to converge to noise-free results when involving complex global illumination. One of the most successful variance-reduction techniques is path guiding, which can learn better distributions for importance sampling to reduce pixel noise. However, previous methods require a large number of path samples to achieve reliable path guiding. We present a novel neural path guiding approach that can reconstruct high-quality sampling distributions for path guiding from a sparse set of samples, using an offline trained neural network. We leverage photons traced from light sources as the primary input for sampling density reconstruction, which is effective for challenging scenes with strong global illumination. To fully make use of our deep neural network, we partition the scene space into an adaptive hierarchical grid, in which we apply our network to reconstruct high-quality sampling distributions for any local region in the scene. This allows for effective path guiding for arbitrary path bounce at any location in path tracing. We demonstrate that our photon-driven neural path guiding approach can generalize to diverse testing scenes, often achieving better rendering results than previous path guiding approaches and opening up interesting future directions.

We propose a novel framework for creating large-scale photorealistic datasets of indoor scenes, with ground truth geometry, material, lighting and semantics. Our goal is to make the dataset creation process widely accessible, transforming scans into photorealistic datasets with high-quality ground truth for appearance, layout, semantic labels, high quality spatially-varying BRDF and complex lighting, including direct, indirect and visibility components. This enables important applications in inverse rendering, scene understanding and robotics. We show that deep networks trained on the proposed dataset achieve competitive performance for shape, material and lighting estimation on real images, enabling photorealistic augmented reality applications, such as object insertion and material editing. We also show our semantic labels may be used for segmentation and multi-task learning. Finally, we demonstrate that our framework may also be integrated with physics engines, to create virtual robotics environments with unique ground truth such as friction coefficients and correspondence to real scenes. The dataset and all the tools to create such datasets will be made publicly available.1

Recent work [28], [5] has demonstrated that volumetric scene representations combined with differentiable volume rendering can enable photo-realistic rendering for challenging scenes that mesh reconstruction fails on. However, these methods entangle geometry and appearance in a "black-box" volume that cannot be edited. Instead, we present an approach that explicitly disentangles geometry—represented as a continuous 3D volume—from appearance—represented as a continuous 2D texture map. We achieve this by introducing a 3D-to-2D texture mapping (or surface parameterization) network into volumetric representations. We constrain this texture mapping network using an additional 2D-to-3D inverse mapping network and a novel cycle consistency loss to make 3D surface points map to 2D texture points that map back to the original 3D points. We demonstrate that this representation can be reconstructed using only multi-view image supervision and generates high-quality rendering results. More importantly, by separating geometry and texture, we allow users to edit appearance by simply editing 2D texture maps.

We present MVSNeRF, a novel neural rendering approach that can efficiently reconstruct neural radiance fields for view synthesis. Unlike prior works on neural radiance fields that consider per-scene optimization on densely captured images, we propose a generic deep neural network that can reconstruct radiance fields from only three nearby input views via fast network inference. Our approach leverages plane-swept cost volumes (widely used in multi-view stereo) for geometry-aware scene reasoning, and combines this with physically based volume rendering for neural radiance field reconstruction. We train our network on real objects in the DTU dataset, and test it on three different datasets to evaluate its effectiveness and generalizability. Our approach can generalize across scenes (even indoor scenes, completely different from our training scenes of objects) and generate realistic view synthesis results using only three input images, significantly outperforming concurrent works on generalizable radiance field reconstruction. Moreover, if dense images are captured, our estimated radiance field representation can be easily fine-tuned; this leads to fast per-scene reconstruction with higher rendering quality and substantially less optimization time than NeRF.

To render higher quality images from the samples generated by path tracing with a low sample count, we propose a novel path reuse approach that processes a fixed collection of paths to iteratively refine and improve radiance estimates throughout the scene. Our method operates on a path graph consisting of the union of the traced paths with additional neighbor edges inserted among clustered nearby vertices. Our approach refines the initial noisy radiance estimates via an aggregation operator, treating vertices within clusters as independent sampling techniques that can be combined using MIS. In a novel step, we also introduce a propagation operator to forward the refined estimates along the paths to successive bounces. We apply the aggregation and propagation operations to the graph iteratively, progressively refining the radiance values, converging to fixed-point radiance estimates with lower variance than the original ones. We also introduce a decorrelation (final gather) step, which uses information already in the graph and is cheap to compute, allowing us to combine the method with standard denoisers. Our approach is lightweight, in the sense that it can be easily plugged into any standard path tracer and neural final image denoiser. Furthermore, it is independent of scene complexity, as the graph size only depends on image resolution and average path depth. We demonstrate that our technique leads to realistic rendering results starting from as low as 1 path per pixel, even in complex indoor scenes dominated by multi-bounce indirect illumination.

We propose NeuMIP, a neural method for representing and rendering a variety of material appearances at different scales. Classical prefiltering (mipmapping) methods work well on simple material properties such as diffuse color, but fail to generalize to normals, self-shadowing, fibers or more complex microstructures and reflectances. In this work, we generalize traditional mipmap pyramids to pyramids of neural textures, combined with a fully connected network. We also introduce neural offsets, a novel method which enables rendering materials with intricate parallax effects without any tessellation. This generalizes classical parallax mapping, but is trained without supervision by any explicit heightfield. Neural materials within our system support a 7-dimensional query, including position, incoming and outgoing direction, and the desired filter kernel size. The materials have small storage (on the order of standard mipmapping except with more texture channels), and can be integrated within common Monte-Carlo path tracing systems. We demonstrate our method on a variety of materials, resulting in complex appearance across levels of detail, with accurate parallax, self-shadowing, and other effects.

Path guiding is a promising technique to reduce the variance of path tracing. Although existing online path guiding algorithms can eventually learn good sampling distributions given a large amount of time and samples, the speed of learning becomes a major bottleneck. In this paper, we accelerate the learning of sampling distributions by training a light-weight neural network offline to reconstruct from sparse samples. Uniquely, we design our neural network to directly operate convolutions on a sparse quadtree, which regresses a high-quality hierarchical sampling distribution. Our approach can reconstruct reasonably accurate sampling distributions faster, allowing for efficient path guiding and rendering. In contrast to the recent offline neural path guiding techniques that reconstruct low-resolution 2D images for sampling, our novel hierarchical framework enables more fine-grained directional sampling with less memory usage, effectively advancing the practicality and efficiency of neural path guiding. In addition, we take advantage of hybrid bidirectional samples including both path samples and photons, as we have found this more robust to different light transport scenarios compared to using only one type of sample as in previous work. Experiments on diverse testing scenes demonstrate that our approach often improves rendering results with better visual quality and lower errors. Our framework can also provide the proper balance of speed, memory cost, and robustness.

We introduce a novel learning-based method to reconstruct the high-quality geometry and complex, spatially-varying BRDF of an arbitrary object from a sparse set of only six images captured by wide-baseline cameras under collocated point lighting. We first estimate per-view depth maps using a deep multi-view stereo network; these depth maps are used to coarsely align the different views. We propose a novel multi-view reflectance estimation network architecture that is trained to pool features from these coarsely aligned images and predict per-view spatially-varying diffuse albedo, surface normals, specular roughness and specular albedo. We do this by jointly optimizing the latent space of our multi-view reflectance network to minimize the photometric error between images rendered with our predictions and the input images. While previous state-of-the-art methods fail on such sparse acquisition setups, we demonstrate, via extensive experiments on synthetic and real data, that our method produces high-quality reconstructions that can be used to render photorealistic images.

We present Uncertainty-aware Cascaded Stereo Network (UCS-Net) for 3D reconstruction from multiple RGB images. Multi-view stereo (MVS) aims to reconstruct fine-grained scene geometry from multi-view images. Previous learning-based MVS methods estimate per-view depth using plane sweep volumes (PSVs) with a fixed depth hypothesis at each plane; this requires densely sampled planes for high accuracy, which is impractical for high-resolution depth because of limited memory. In contrast, we propose adaptive thin volumes (ATVs); in an ATV, the depth hypothesis of each plane is spatially varying, which adapts to the uncertainties of previous per-pixel depth predictions. Our UCS-Net has three stages: the first stage processes a small PSV to predict low-resolution depth; two ATVs are then used in the following stages to refine the depth with higher resolution and higher accuracy. Our ATV consists of only a small number of planes with low memory and computation costs; yet, it efficiently partitions local depth ranges within learned small uncertainty intervals. We propose to use variance-based uncertainty estimates to adaptively construct ATVs; this differentiable process leads to reasonable and fine-grained spatial partitioning. Our multi-stage framework progressively sub-divides the vast scene space with increasing depth resolution and precision, which enables reconstruction with high completeness and accuracy in a coarse-to-fine fashion. We demonstrate that our method achieves superior performance compared with other learning-based MVS methods on various challenging datasets.

Abstract Recently, deep learning‐based denoising approaches have led to dramatic improvements in low sample‐count Monte Carlo rendering. These approaches are aimed at path tracing, which is not ideal for simulating challenging light transport effects like caustics, where photon mapping is the method of choice. However, photon mapping requires very large numbers of traced photons to achieve high‐quality reconstructions. In this paper, we develop the first deep learning‐based method for particle‐based rendering, and specifically focus on photon density estimation, the core of all particle‐based methods. We train a novel deep neural network to predict a kernel function to aggregate photon contributions at shading points. Our network encodes individual photons into per‐photon features, aggregates them in the neighborhood of a shading point to construct a photon local context vector, and infers a kernel function from the per‐photon and photon local context features. This network is easy to incorporate in many previous photon mapping methods (by simply swapping the kernel density estimator) and can produce high‐quality reconstructions of complex global illumination effects like caustics with an order of magnitude fewer photons compared to previous photon mapping methods. Our approach largely reduces the required number of photons, significantly advancing the computational efficiency in photon mapping.

The light stage has been widely used in computer graphics for the past two decades, primarily to enable the relighting of human faces. By capturing the appearance of the human subject under different light sources, one obtains the light transport matrix of that subject, which enables image-based relighting in novel environments. However, due to the finite number of lights in the stage, the light transport matrix only represents a sparse sampling on the entire sphere. As a consequence, relighting the subject with a point light or a directional source that does not coincide exactly with one of the lights in the stage requires interpolation and resampling the images corresponding to nearby lights, and this leads to ghosting shadows, aliased specularities, and other artifacts. To ameliorate these artifacts and produce better results under arbitrary high-frequency lighting, this paper proposes a learning-based solution for the "super-resolution" of scans of human faces taken from a light stage. Given an arbitrary "query" light direction, our method aggregates the captured images corresponding to neighboring lights in the stage, and uses a neural network to synthesize a rendering of the face that appears to be illuminated by a "virtual" light source at the query location. This neural network must circumvent the inherent aliasing and regularity of the light stage data that was used for training, which we accomplish through the use of regularized traditional interpolation methods within our network. Our learned model is able to produce renderings for arbitrary light directions that exhibit realistic shadows and specular highlights, and is able to generalize across a wide variety of subjects. Our super-resolution approach enables more accurate renderings of human subjects under detailed environment maps, or the construction of simpler light stages that contain fewer light sources while still yielding comparable quality renderings as light stages with more densely sampled lights.

Rendering specular material appearance is a core problem of computer graphics. While smooth analytical material models are widely used, the high-frequency structure of real specular highlights requires considering discrete, finite microgeometry. Instead of explicit modeling and simulation of the surface microstructure (which was explored in previous work), we propose a novel direction: learning the high-frequency directional patterns from synthetic or measured examples, by training a generative adversarial network (GAN). A key challenge in applying GAN synthesis to spatially varying BRDFs is evaluating the reflectance for a single location and direction without the cost of evaluating the whole hemisphere. We resolve this using a novel method for partial evaluation of the generator network. We are also able to control large-scale spatial texture using a conditional GAN approach. The benefits of our approach include the ability to synthesize spatially large results without repetition, support for learning from measured data, and evaluation performance independent of the complexity of the dataset synthesis or measurement.

Lighting plays a central role in conveying the essence and depth of the subject in a portrait photograph. Professional photographers will carefully control the lighting in their studio to manipulate the appearance of their subject, while consumer photographers are usually constrained to the illumination of their environment. Though prior works have explored techniques for relighting an image, their utility is usually limited due to requirements of specialized hardware, multiple images of the subject under controlled or known illuminations, or accurate models of geometry and reflectance. To this end, we present a system for portrait relighting : a neural network that takes as input a single RGB image of a portrait taken with a standard cellphone camera in an unconstrained environment, and from that image produces a relit image of that subject as though it were illuminated according to any provided environment map. Our method is trained on a small database of 18 individuals captured under different directional light sources in a controlled light stage setup consisting of a densely sampled sphere of lights. Our proposed technique produces quantitatively superior results on our dataset's validation set compared to prior works, and produces convincing qualitative relighting results on a dataset of hundreds of real-world cellphone portraits. Because our technique can produce a 640 × 640 image in only 160 milliseconds, it may enable interactive user-facing photographic applications in the future.

The goal of light transport acquisition is to take images from a sparse set of lighting and viewing directions, and combine them to enable arbitrary relighting with changing view. While relighting from sparse images has received significant attention, there has been relatively less progress on view synthesis from a sparse set of "photometric" images---images captured under controlled conditions, lit by a single directional source; we use a spherical gantry to position the camera on a sphere surrounding the object. In this paper, we synthesize novel viewpoints across a wide range of viewing directions (covering a 60° cone) from a sparse set of just six viewing directions. While our approach relates to previous view synthesis and image-based rendering techniques, those methods are usually restricted to much smaller baselines, and are captured under environment illumination. At our baselines, input images have few correspondences and large occlusions; however we benefit from structured photometric images. Our method is based on a deep convolutional network trained to directly synthesize new views from the six input views. This network combines 3D convolutions on a plane sweep volume with a novel per-view per-depth plane attention map prediction network to effectively aggregate multi-view appearance. We train our network with a large-scale synthetic dataset of 1000 scenes with complex geometry and material properties. In practice, it is able to synthesize novel viewpoints for captured real data and reproduces complex appearance effects like occlusions, view-dependent specularities and hard shadows. Moreover, the method can also be combined with previous relighting techniques to enable changing both lighting and view, and applied to computer vision problems like multiview stereo from sparse image sets.

Reconstructing shape and reflectance properties from images is a highly under-constrained problem, and has previously been addressed by using specialized hardware to capture calibrated data or by assuming known (or highly constrained) shape or reflectance. In contrast, we demonstrate that we can recover non-Lambertian, spatially-varying BRDFs and complex geometry belonging to any arbitrary shape class, from a single RGB image captured under a combination of unknown environment illumination and flash lighting. We achieve this by training a deep neural network to regress shape and reflectance from the image. Our network is able to address this problem because of three novel contributions: first, we build a large-scale dataset of procedurally generated shapes and real-world complex SVBRDFs that approximate real world appearance well. Second, single image inverse rendering requires reasoning at multiple scales, and we propose a cascade network structure that allows this in a tractable manner. Finally, we incorporate an in-network rendering layer that aids the reconstruction task by handling global illumination effects that are important for real-world scenes. Together, these contributions allow us to tackle the entire inverse rendering problem in a holistic manner and produce state-of-the-art results on both synthetic and real data.

We present an image-based relighting method that can synthesize scene appearance under novel, distant illumination from the visible hemisphere, from only five images captured under pre-defined directional lights. Our method uses a deep convolutional neural network to regress the relit image from these five images; this relighting network is trained on a large synthetic dataset comprised of procedurally generated shapes with real-world reflectances. We show that by combining a custom-designed sampling network with the relighting network, we can jointly learn both the optimal input light directions and the relighting function. We present an extensive evaluation of our network, including an empirical analysis of reconstruction quality, optimal lighting configurations for different scenarios, and alternative network architectures. We demonstrate, on both synthetic and real scenes, that our method is able to reproduce complex, high-frequency lighting effects like specularities and cast shadows, and outperforms other image-based relighting methods that require an order of magnitude more images.

Highly effective optimization frameworks have been developed for traditional multiview stereo relying on lambertian photoconsistency. However, they do not account for complex material properties. On the other hand, recent works have explored PDE invariants for shape recovery with complex BRDFs, but they have not been incorporated into robust numerical optimization frameworks. We present a variational energy minimization framework for robust recovery of shape in multiview stereo with complex, unknown BRDFs. While our formulation is general, we demonstrate its efficacy on shape recovery using a single light field image, where the microlens array may be considered as a realization of a purely translational multiview stereo setup. Our formulation automatically balances contributions from texture gradients, traditional Lambertian photoconsistency, an appropriate BRDF-invariant PDE and a smoothness prior. Unlike prior works, our energy function inherently handles spatially-varying BRDFs and albedos. Extensive experiments with synthetic and real data show that our optimization framework consistently achieves errors lower than Lambertian baselines and further, is more robust than prior BRDF-invariant reconstruction methods.

We develop a method to acquire the BRDF of a homogeneous flat sample from only two images, taken by a near-field perspective camera, and lit by a directional light source. Our method uses the MERL BRDF database to determine the optimal set of lightview pairs for data-driven reflectance acquisition. We develop a mathematical framework to estimate error from a given set of measurements, including the use of multiple measurements in an image simultaneously, as needed for acquisition from near-field setups. The novel error metric is essential in the near-field case, where we show that using the condition-number alone performs poorly. We demonstrate practical near-field acquisition of BRDFs from only one or two input images. Our framework generalizes to configurations like a fixed camera setup, where we also develop a simple extension to spatially-varying BRDFs by clustering the materials.

Dynamic hair strands have complex structures and experience intricate collisions and occlusion, posing significant challenges for high-quality reconstruction of their motions. We present a comprehensive dynamic hair capture system for reconstructing realistic hair motions from multiple synchronized video sequences. To recover hair strands' temporal correspondence, we propose a motion-path analysis algorithm that can robustly track local hair motions in input videos. To ensure the spatial and temporal coherence of the dynamic capture, we formulate the global hair reconstruction as a spacetime optimization problem solved iteratively. Demonstrated using a range of real-world hairstyles driven by different wind conditions and head motions, our approach is able to reconstruct complex hair dynamics matching closely with video recordings both in terms of geometry and motion details.