Sai Bi

Kai Zhang, Nick Kolkin, Sai Bi et al. · deepmind

We present a method for transferring the artistic features of an arbitrary style image to a 3D scene. Previous methods that perform 3D stylization on point clouds or meshes are sensitive to geometric reconstruction errors for complex real-world scenes. Instead, we propose to stylize the more robust radiance field representation. We find that the commonly used Gram matrix-based loss tends to produce blurry results without faithful brushstrokes, and introduce a nearest neighbor-based loss that is highly effective at capturing style details while maintaining multi-view consistency. We also propose a novel deferred back-propagation method to enable optimization of memory-intensive radiance fields using style losses defined on full-resolution rendered images. Our extensive evaluation demonstrates that our method outperforms baselines by generating artistic appearance that more closely resembles the style image. Please check our project page for video results and open-source implementations: https://www.cs.cornell.edu/projects/arf/ .

Sai Praveen Bangaru, Michaël Gharbi, Tzu-Mao Li et al.

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.

Zhengqin Li, Jia Shi, Sai Bi et al.

We present a method to edit complex indoor lighting from a single image with its predicted depth and light source segmentation masks. This is an extremely challenging problem that requires modeling complex light transport, and disentangling HDR lighting from material and geometry with only a partial LDR observation of the scene. We tackle this problem using two novel components: 1) a holistic scene reconstruction method that estimates scene reflectance and parametric 3D lighting, and 2) a neural rendering framework that re-renders the scene from our predictions. We use physically-based indoor light representations that allow for intuitive editing, and infer both visible and invisible light sources. Our neural rendering framework combines physically-based direct illumination and shadow rendering with deep networks to approximate global illumination. It can capture challenging lighting effects, such as soft shadows, directional lighting, specular materials, and interreflections. Previous single image inverse rendering methods usually entangle scene lighting and geometry and only support applications like object insertion. Instead, by combining parametric 3D lighting estimation with neural scene rendering, we demonstrate the first automatic method to achieve full scene relighting, including light source insertion, removal, and replacement, from a single image. All source code and data will be publicly released.

Haian Jin, Isabella Liu, Peijia Xu et al.

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.

Jiahao Li, Hao Tan, Kai Zhang et al.

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: https://jiahao.ai/instant3d/.

Zhengfei Kuang, Fujun Luan, Sai Bi et al.

Recent advances in neural radiance fields have enabled the high-fidelity 3D reconstruction of complex scenes for novel view synthesis. However, it remains underexplored how the appearance of such representations can be efficiently edited while maintaining photorealism. In this work, we present PaletteNeRF, a novel method for photorealistic appearance editing of neural radiance fields (NeRF) based on 3D color decomposition. Our method decomposes the appearance of each 3D point into a linear combination of palette-based bases (i.e., 3D segmentations defined by a group of NeRF-type functions) that are shared across the scene. While our palette-based bases are view-independent, we also predict a view-dependent function to capture the color residual (e.g., specular shading). During training, we jointly optimize the basis functions and the color palettes, and we also introduce novel regularizers to encourage the spatial coherence of the decomposition. Our method allows users to efficiently edit the appearance of the 3D scene by modifying the color palettes. We also extend our framework with compressed semantic features for semantic-aware appearance editing. We demonstrate that our technique is superior to baseline methods both quantitatively and qualitatively for appearance editing of complex real-world scenes.

Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli et al.

While NeRF 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 NeRFusion achieves 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.

Yinghao Xu, Hao Tan, Fujun Luan et al.

We propose \textbf{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 can denoise noisy multi-view images via NeRF reconstruction and rendering, achieving single-stage 3D generation in $\sim$30s on single A100 GPU. We train \textbf{DMV3D} on large-scale multi-view image datasets of highly 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://justimyhxu.github.io/projects/dmv3d/ .

Peng Wang, Hao Tan, Sai Bi et al.

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 the 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 .

Yicong Hong, Kai Zhang, Jiuxiang Gu et al.

We propose the first Large Reconstruction Model (LRM) that predicts the 3D model of an object from a single input image within just 5 seconds. In contrast to many previous methods that are trained on small-scale datasets such as ShapeNet in a category-specific fashion, LRM adopts a highly scalable transformer-based architecture with 500 million learnable parameters to directly predict a neural radiance field (NeRF) from the input image. We train our model in an end-to-end manner on massive multi-view data containing around 1 million objects, including both synthetic renderings from Objaverse and real captures from MVImgNet. This combination of a high-capacity model and large-scale training data empowers our model to be highly generalizable and produce high-quality 3D reconstructions from various testing inputs, including real-world in-the-wild captures and images created by generative models. Video demos and interactable 3D meshes can be found on our LRM project webpage: https://yiconghong.me/LRM.

Nithin Raghavan, Yan Xiao, Kai-En Lin et al.

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.

Guangyan Cai, Fujun Luan, Miloš Hašan et al.

Glossy objects present a significant challenge for 3D reconstruction from multi-view input images under natural lighting. In this paper, we introduce PBIR-NIE, an inverse rendering framework designed to holistically capture the geometry, material attributes, and surrounding illumination of such objects. We propose a novel parallax-aware non-distant environment map as a lightweight and efficient lighting representation, accurately modeling the near-field background of the scene, which is commonly encountered in real-world capture setups. This feature allows our framework to accommodate complex parallax effects beyond the capabilities of standard infinite-distance environment maps. Our method optimizes an underlying signed distance field (SDF) through physics-based differentiable rendering, seamlessly connecting surface gradients between a triangle mesh and the SDF via neural implicit evolution (NIE). To address the intricacies of highly glossy BRDFs in differentiable rendering, we integrate the antithetic sampling algorithm to mitigate variance in the Monte Carlo gradient estimator. Consequently, our framework exhibits robust capabilities in handling glossy object reconstruction, showcasing superior quality in geometry, relighting, and material estimation.

Qitao Zhao, Hao Tan, Qianqian Wang et al.

Self-supervised pre-training has revolutionized foundation models for languages, individual 2D images and videos, but remains largely unexplored for learning 3D-aware representations from multi-view images. In this paper, we present E-RayZer, a self-supervised large 3D Vision model that learns truly 3D-aware representations directly from unlabeled images. Unlike prior self-supervised methods such as RayZer that infer 3D indirectly through latent-space view synthesis, E-RayZer operates directly in 3D space, performing self-supervised 3D reconstruction with Explicit geometry. This formulation eliminates shortcut solutions and yields representations that are geometrically grounded. To ensure convergence and scalability, we introduce a novel fine-grained learning curriculum that organizes training from easy to hard samples and harmonizes heterogeneous data sources in an entirely unsupervised manner. Experiments demonstrate that E-RayZer significantly outperforms RayZer on pose estimation, matches or sometimes surpasses fully supervised reconstruction models such as VGGT. Furthermore, its learned representations outperform leading visual pre-training models (e.g., DINOv3, CroCo v2, VideoMAE V2, and RayZer) when transferring to 3D downstream tasks, establishing E-RayZer as a new paradigm for 3D-aware visual pre-training.

ShahRukh Athar, Zhixin Shu, Zexiang Xu et al.

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 artifacts 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. The project page can be found here: http://shahrukhathar.github.io/2023/08/22/CoDyNeRF.html

Chen Wang, Hao Tan, Wang Yifan et al.

We propose tttLRM, a novel large 3D reconstruction model that leverages a Test-Time Training (TTT) layer to enable long-context, autoregressive 3D reconstruction with linear computational complexity, further scaling the model's capability. Our framework efficiently compresses multiple image observations into the fast weights of the TTT layer, forming an implicit 3D representation in the latent space that can be decoded into various explicit formats, such as Gaussian Splats (GS) for downstream applications. The online learning variant of our model supports progressive 3D reconstruction and refinement from streaming observations. We demonstrate that pretraining on novel view synthesis tasks effectively transfers to explicit 3D modeling, resulting in improved reconstruction quality and faster convergence. Extensive experiments show that our method achieves superior performance in feedforward 3D Gaussian reconstruction compared to state-of-the-art approaches on both objects and scenes.

Yicong Hong, Yiqun Mei, Chongjian Ge et al.

A truly interactive world model requires three key ingredients: real-time long-horizon streaming, consistent spatial memory, and precise user control. However, most existing approaches address only one of these aspects in isolation, as achieving all three simultaneously is highly challenging-for example, long-term memory mechanisms often degrade real-time performance. In this work, we present RELIC, a unified framework that tackles these three challenges altogether. Given a single image and a text description, RELIC enables memory-aware, long-duration exploration of arbitrary scenes in real time. Built upon recent autoregressive video-diffusion distillation techniques, our model represents long-horizon memory using highly compressed historical latent tokens encoded with both relative actions and absolute camera poses within the KV cache. This compact, camera-aware memory structure supports implicit 3D-consistent content retrieval and enforces long-term coherence with minimal computational overhead. In parallel, we fine-tune a bidirectional teacher video model to generate sequences beyond its original 5-second training horizon, and transform it into a causal student generator using a new memory-efficient self-forcing paradigm that enables full-context distillation over long-duration teacher as well as long student self-rollouts. Implemented as a 14B-parameter model and trained on a curated Unreal Engine-rendered dataset, RELIC achieves real-time generation at 16 FPS while demonstrating more accurate action following, more stable long-horizon streaming, and more robust spatial-memory retrieval compared with prior work. These capabilities establish RELIC as a strong foundation for the next generation of interactive world modeling.

Desai Xie, Jiahao Li, Hao Tan et al.

Multi-view diffusion models, obtained by applying Supervised Finetuning (SFT) to text-to-image diffusion models, have driven recent breakthroughs in text-to-3D research. However, due to the limited size and quality of existing 3D datasets, they still suffer from multi-view inconsistencies and Neural Radiance Field (NeRF) reconstruction artifacts. We argue that multi-view diffusion models can benefit from further Reinforcement Learning Finetuning (RLFT), which allows models to learn from the data generated by themselves and improve beyond their dataset limitations during SFT. To this end, we introduce Carve3D, an improved RLFT algorithm coupled with a novel Multi-view Reconstruction Consistency (MRC) metric, to enhance the consistency of multi-view diffusion models. To measure the MRC metric on a set of multi-view images, we compare them with their corresponding NeRF renderings at the same camera viewpoints. The resulting model, which we denote as Carve3DM, demonstrates superior multi-view consistency and NeRF reconstruction quality than existing models. Our results suggest that pairing SFT with Carve3D's RLFT is essential for developing multi-view-consistent diffusion models, mirroring the standard Large Language Model (LLM) alignment pipeline. Our code, training and testing data, and video results are available at: https://desaixie.github.io/carve-3d.

Kai Zhang, Sai Bi, Hao Tan et al.

We propose GS-LRM, a scalable large reconstruction model that can predict high-quality 3D Gaussian primitives from 2-4 posed sparse images in 0.23 seconds on single A100 GPU. Our model features a very simple transformer-based architecture; we patchify input posed images, pass the concatenated multi-view image tokens through a sequence of transformer blocks, and decode final per-pixel Gaussian parameters directly from these tokens for differentiable rendering. In contrast to previous LRMs that can only reconstruct objects, by predicting per-pixel Gaussians, GS-LRM naturally handles scenes with large variations in scale and complexity. We show that our model can work on both object and scene captures by training it on Objaverse and RealEstate10K respectively. In both scenarios, the models outperform state-of-the-art baselines by a wide margin. We also demonstrate applications of our model in downstream 3D generation tasks. Our project webpage is available at: https://sai-bi.github.io/project/gs-lrm/ .

Xinyue Wei, Kai Zhang, Sai Bi et al.

We propose MeshLRM, a novel LRM-based approach that can reconstruct a high-quality mesh from merely four input images in less than one second. Different from previous large reconstruction models (LRMs) that focus on NeRF-based reconstruction, MeshLRM incorporates differentiable mesh extraction and rendering within the LRM framework. This allows for end-to-end mesh reconstruction by fine-tuning a pre-trained NeRF LRM with mesh rendering. Moreover, we improve the LRM architecture by simplifying several complex designs in previous LRMs. MeshLRM's NeRF initialization is sequentially trained with low- and high-resolution images; this new LRM training strategy enables significantly faster convergence and thereby leads to better quality with less compute. Our approach achieves state-of-the-art mesh reconstruction from sparse-view inputs and also allows for many downstream applications, including text-to-3D and single-image-to-3D generation. Project page: https://sarahweiii.github.io/meshlrm/

Kai Zhang, Peng Wang, Sai Bi et al.

We present KnapFormer, an efficient and versatile framework to combine workload balancing and sequence parallelism in distributed training of Diffusion Transformers (DiT). KnapFormer builds on the insight that strong synergy exists between sequence parallelism and the need to address the significant token imbalance across ranks. This imbalance arises from variable-length text inputs and varying visual token counts in mixed-resolution and image-video joint training. KnapFormer redistributes tokens by first gathering sequence length metadata across all ranks in a balancing group and solving a global knapsack problem. The solver aims to minimize the variances of total workload per-GPU, while accounting for the effect of sequence parallelism. By integrating DeepSpeed-Ulysees-based sequence parallelism in the load-balancing decision process and utilizing a simple semi-empirical workload model, KnapFormers achieves minimal communication overhead and less than 1% workload discrepancy in real-world training workloads with sequence length varying from a few hundred to tens of thousands. It eliminates straggler effects and achieves 2x to 3x speedup when training state-of-the-art diffusion models like FLUX on mixed-resolution and image-video joint data corpora. We open-source the KnapFormer implementation at https://github.com/Kai-46/KnapFormer/

Haian Jin, Hanwen Jiang, Hao Tan et al. · deepmind

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. Notably, our models surpass all previous methods even with reduced computational resources (1-2 GPUs). Please see our website for more details: https://haian-jin.github.io/projects/LVSM/ .

Chen Ziwen, Hao Tan, Kai Zhang et al.

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

Tianyuan Zhang, Sai Bi, Yicong Hong et al.

Test-Time Training (TTT) models context dependencies by adapting part of the model's weights (referred to as fast weights) during inference. This fast weight, akin to recurrent states in RNNs, stores temporary memories of past tokens in the current sequence. Existing TTT methods struggled to show effectiveness in handling long-context data, due to their inefficiency on modern GPUs. The TTT layers in many of these approaches operate with extremely low FLOPs utilization (often <5%) because they deliberately apply small online minibatch sizes (e.g., updating fast weights every 16 or 64 tokens). Moreover, a small minibatch implies fine-grained block-wise causal dependencies in the data, unsuitable for data beyond 1D ordered sequences, like sets or N-dimensional grids such as images or videos. In contrast, we pursue the opposite direction by using an extremely large chunk update, ranging from 2K to 1M tokens across tasks of varying modalities, which we refer to as Large Chunk Test-Time Training (LaCT). It improves hardware utilization by orders of magnitude, and more importantly, facilitates scaling of nonlinear state size (up to 40% of model parameters), hence substantially improving state capacity, all without requiring cumbersome and error-prone kernel implementations. It also allows easy integration of sophisticated optimizers, e.g. Muon for online updates. We validate our approach across diverse modalities and tasks, including novel view synthesis with image set, language models, and auto-regressive video diffusion. Our approach can scale up to 14B-parameter AR video diffusion model on sequences up to 56K tokens. In our longest sequence experiment, we perform novel view synthesis with 1 million context length. We hope this work will inspire and accelerate new research in the field of long-context modeling and test-time training. Website: https://tianyuanzhang.com/projects/ttt-done-right

Liwen Wu, Sai Bi, Zexiang Xu et al.

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 project webpage and source code are available at: \url{https://lwwu2.github.io/nde/}.

Hanwen Jiang, Zexiang Xu, Desai Xie et al.

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.

Hanwen Jiang, Hao Tan, Peng Wang et al.

We present RayZer, a self-supervised multi-view 3D Vision model trained without any 3D supervision, i.e., camera poses and scene geometry, while exhibiting emerging 3D awareness. Concretely, RayZer takes unposed and uncalibrated images as input, recovers camera parameters, reconstructs a scene representation, and synthesizes novel views. During training, RayZer relies solely on its self-predicted camera poses to render target views, eliminating the need for any ground-truth camera annotations and allowing RayZer to be trained with 2D image supervision. The emerging 3D awareness of RayZer is attributed to two key factors. First, we design a self-supervised framework, which achieves 3D-aware auto-encoding of input images by disentangling camera and scene representations. Second, we design a transformer-based model in which the only 3D prior is the ray structure, connecting camera, pixel, and scene simultaneously. RayZer demonstrates comparable or even superior novel view synthesis performance than ``oracle'' methods that rely on pose annotations in both training and testing. Project: https://hwjiang1510.github.io/RayZer/

Hanzhe Hu, Tianwei Yin, Fujun Luan et al.

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. 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 multi-view 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.

Sara Rojas, Julien Philip, Kai Zhang et al.

Recent advancements in diffusion models have shown remarkable proficiency in editing 2D images based on text prompts. However, extending these techniques to edit scenes in Neural Radiance Fields (NeRF) is complex, as editing individual 2D frames can result in inconsistencies across multiple views. Our crucial insight is that a NeRF scene's geometry can serve as a bridge to integrate these 2D edits. Utilizing this geometry, we employ a depth-conditioned ControlNet to enhance the coherence of each 2D image modification. Moreover, we introduce an inpainting approach that leverages the depth information of NeRF scenes to distribute 2D edits across different images, ensuring robustness against errors and resampling challenges. Our results reveal that this methodology achieves more consistent, lifelike, and detailed edits than existing leading methods for text-driven NeRF scene editing.

Hansheng Chen, Kai Zhang, Hao Tan et al.

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.

Zhengfei Kuang, Tianyuan Zhang, Kai Zhang et al.

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.

Gene Chou, Kai Zhang, Sai Bi et al. · deepmind

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.

Ziqiao Ma, Xuweiyi Chen, Shoubin Yu et al.

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.

Hansheng Chen, Kai Zhang, Hao Tan et al.

Few-step diffusion or flow-based generative models typically distill a velocity-predicting teacher into a student that predicts a shortcut towards denoised data. This format mismatch has led to complex distillation procedures that often suffer from a quality-diversity trade-off. To address this, we propose policy-based flow models ($π$-Flow). $π$-Flow modifies the output layer of a student flow model to predict a network-free policy at one timestep. The policy then produces dynamic flow velocities at future substeps with negligible overhead, enabling fast and accurate ODE integration on these substeps without extra network evaluations. To match the policy's ODE trajectory to the teacher's, we introduce a novel imitation distillation approach, which matches the policy's velocity to the teacher's along the policy's trajectory using a standard $\ell_2$ flow matching loss. By simply mimicking the teacher's behavior, $π$-Flow enables stable and scalable training and avoids the quality-diversity trade-off. On ImageNet 256$^2$, it attains a 1-NFE FID of 2.85, outperforming MeanFlow of the same DiT architecture. On FLUX.1-12B and Qwen-Image-20B at 4 NFEs, $π$-Flow achieves substantially better diversity than state-of-the-art few-step methods, while maintaining teacher-level quality.

Bowei Chen, Sai Bi, Hao Tan et al.

In this work, we propose aligning pretrained visual encoders to serve as tokenizers for latent diffusion models in image generation. Unlike training a variational autoencoder (VAE) from scratch, which primarily emphasizes low-level details, our approach leverages the rich semantic structure of foundation encoders. We introduce a three-stage alignment strategy: (1) freeze the encoder and train an adapter and a decoder to establish a semantic latent space; (2) jointly optimize all components with an additional semantic preservation loss, enabling the encoder to capture perceptual details while retaining high-level semantics; and (3) refine the decoder for improved reconstruction quality. This alignment yields semantically rich image tokenizers that benefit diffusion models. On ImageNet 256$\times$256, our tokenizer accelerates the convergence of diffusion models, reaching a gFID of 1.90 within just 64 epochs, and improves generation both with and without classifier-free guidance. Scaling to LAION, a 2B-parameter text-to-image model trained with our tokenizer consistently outperforms FLUX VAE under the same training steps. Overall, our method is simple, scalable, and establishes a semantically grounded paradigm for continuous tokenizer design.

Kai Zhang, Fujun Luan, Sai Bi et al.

Classifier-free guidance (CFG) is widely used in diffusion models but often introduces over-contrast and over-saturation artifacts at higher guidance strengths. We present EP-CFG (Energy-Preserving Classifier-Free Guidance), which addresses these issues by preserving the energy distribution of the conditional prediction during the guidance process. Our method simply rescales the energy of the guided output to match that of the conditional prediction at each denoising step, with an optional robust variant for improved artifact suppression. Through experiments, we show that EP-CFG maintains natural image quality and preserves details across guidance strengths while retaining CFG's semantic alignment benefits, all with minimal computational overhead.

Liwen Wu, Sai Bi, Zexiang Xu et al.

Neural bidirectional reflectance distribution functions (BRDFs) have emerged as popular material representations for enhancing realism in physically-based rendering. Yet their importance sampling remains a significant challenge. In this paper, we introduce a reparameterization-based formulation of neural BRDF importance sampling that seamlessly integrates into the standard rendering pipeline with precise generation of BRDF samples. The reparameterization-based formulation transfers the distribution learning task to a problem of identifying BRDF integral substitutions. In contrast to previous methods that rely on invertible networks and multi-step inference to reconstruct BRDF distributions, our model removes these constraints, which offers greater flexibility and efficiency. Our variance and performance analysis demonstrates that our reparameterization method achieves the best variance reduction in neural BRDF renderings while maintaining high inference speeds compared to existing baselines.

Desai Xie, Sai Bi, Zhixin Shu et al.

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/.

Haian Jin, Yuan Li, Fujun Luan et al.

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.

Yiqun Mei, Yu Zeng, He Zhang et al.

At the core of portrait photography is the search for ideal lighting and viewpoint. The process often requires advanced knowledge in photography and an elaborate studio setup. In this work, we propose Holo-Relighting, a volumetric relighting method that is capable of synthesizing novel viewpoints, and novel lighting from a single image. Holo-Relighting leverages the pretrained 3D GAN (EG3D) to reconstruct geometry and appearance from an input portrait as a set of 3D-aware features. We design a relighting module conditioned on a given lighting to process these features, and predict a relit 3D representation in the form of a tri-plane, which can render to an arbitrary viewpoint through volume rendering. Besides viewpoint and lighting control, Holo-Relighting also takes the head pose as a condition to enable head-pose-dependent lighting effects. With these novel designs, Holo-Relighting can generate complex non-Lambertian lighting effects (e.g., specular highlights and cast shadows) without using any explicit physical lighting priors. We train Holo-Relighting with data captured with a light stage, and propose two data-rendering techniques to improve the data quality for training the volumetric relighting system. Through quantitative and qualitative experiments, we demonstrate Holo-Relighting can achieve state-of-the-arts relighting quality with better photorealism, 3D consistency and controllability.

Xinyue Wei, Fanbo Xiang, Sai Bi et al.

We present a method for generating high-quality watertight manifold meshes from multi-view input images. Existing volumetric rendering methods are robust in optimization but tend to generate noisy meshes with poor topology. Differentiable rasterization-based methods can generate high-quality meshes but are sensitive to initialization. Our method combines the benefits of both worlds; we take the geometry initialization obtained from neural volumetric fields, and further optimize the geometry as well as a compact neural texture representation with differentiable rasterizers. Through extensive experiments, we demonstrate that our method can generate accurate mesh reconstructions with faithful appearance that are comparable to previous volume rendering methods while being an order of magnitude faster in rendering. We also show that our generated mesh and neural texture reconstruction is compatible with existing graphics pipelines and enables downstream 3D applications such as simulation. Project page: https://sarahweiii.github.io/neumanifold/

Qiangeng Xu, Zexiang Xu, Julien Philip et al.

Volumetric neural rendering methods like NeRF 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 30X faster training time. Point-NeRF can be combined with other 3D reconstruction methods and handles the errors and outliers in such methods via a novel pruning and growing mechanism. The experiments on the DTU, the NeRF Synthetics , the ScanNet and the Tanks and Temples datasets demonstrate Point-NeRF can surpass the existing methods and achieve the state-of-the-art results.

Mohammad Shafiei, Sai Bi, Zhengqin Li et al.

Recently neural volumetric representations such as neural reflectance fields have been widely applied to faithfully reproduce the appearance of real-world objects and scenes under novel viewpoints and lighting conditions. However, it remains challenging and time-consuming to render such representations under complex lighting such as environment maps, which requires individual ray marching towards each single light to calculate the transmittance at every sampled point. In this paper, we propose a novel method based on precomputed Neural Transmittance Functions to accelerate the rendering of neural reflectance fields. Our neural transmittance functions enable us to efficiently query the transmittance at an arbitrary point in space along an arbitrary ray without tedious ray marching, which effectively reduces the time-complexity of the rendering. We propose a novel formulation for the neural transmittance function, and train it jointly with the neural reflectance fields on images captured under collocated camera and light, while enforcing monotonicity. Results on real and synthetic scenes demonstrate almost two order of magnitude speedup for renderings under environment maps with minimal accuracy loss.

Tiancheng Sun, Kai-En Lin, Sai Bi et al.

Human portraits exhibit various appearances when observed from different views under different lighting conditions. We can easily imagine how the face will look like in another setup, but computer algorithms still fail on this problem given limited observations. To this end, we present a system for portrait view synthesis and relighting: given multiple portraits, we use a neural network to predict the light-transport field in 3D space, and from the predicted Neural Light-transport Field (NeLF) produce a portrait from a new camera view under a new environmental lighting. Our system is trained on a large number of synthetic models, and can generalize to different synthetic and real portraits under various lighting conditions. Our method achieves simultaneous view synthesis and relighting given multi-view portraits as the input, and achieves state-of-the-art results.

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.

Zhengqin Li, Ting-Wei Yu, Shen Sang et al.

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.

Sai Bi, Zexiang Xu, Kalyan Sunkavalli et al.

We present a deep learning approach to reconstruct scene appearance from unstructured images captured under collocated point lighting. At the heart of Deep Reflectance Volumes is a novel volumetric scene representation consisting of opacity, surface normal and reflectance voxel grids. We present a novel physically-based differentiable volume ray marching framework to render these scene volumes under arbitrary viewpoint and lighting. This allows us to optimize the scene volumes to minimize the error between their rendered images and the captured images. Our method is able to reconstruct real scenes with challenging non-Lambertian reflectance and complex geometry with occlusions and shadowing. Moreover, it accurately generalizes to novel viewpoints and lighting, including non-collocated lighting, rendering photorealistic images that are significantly better than state-of-the-art mesh-based methods. We also show that our learned reflectance volumes are editable, allowing for modifying the materials of the captured scenes.

Sai Bi, Kalyan Sunkavalli, Federico Perazzi et al.

We present a method to improve the visual realism of low-quality, synthetic images, e.g. OpenGL renderings. Training an unpaired synthetic-to-real translation network in image space is severely under-constrained and produces visible artifacts. Instead, we propose a semi-supervised approach that operates on the disentangled shading and albedo layers of the image. Our two-stage pipeline first learns to predict accurate shading in a supervised fashion using physically-based renderings as targets, and further increases the realism of the textures and shading with an improved CycleGAN network. Extensive evaluations on the SUNCG indoor scene dataset demonstrate that our approach yields more realistic images compared to other state-of-the-art approaches. Furthermore, networks trained on our generated "real" images predict more accurate depth and normals than domain adaptation approaches, suggesting that improving the visual realism of the images can be more effective than imposing task-specific losses.

Sai Bi, Zexiang Xu, Kalyan Sunkavalli et al.

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.

Sai Bi, Nima Khademi Kalantari, Ravi Ramamoorthi

Intrinsic image decomposition is the process of separating the reflectance and shading layers of an image, which is a challenging and underdetermined problem. In this paper, we propose to systematically address this problem using a deep convolutional neural network (CNN). Although deep learning (DL) has been recently used to handle this application, the current DL methods train the network only on synthetic images as obtaining ground truth reflectance and shading for real images is difficult. Therefore, these methods fail to produce reasonable results on real images and often perform worse than the non-DL techniques. We overcome this limitation by proposing a novel hybrid approach to train our network on both synthetic and real images. Specifically, in addition to directly supervising the network using synthetic images, we train the network by enforcing it to produce the same reflectance for a pair of images of the same real-world scene with different illuminations. Furthermore, we improve the results by incorporating a bilateral solver layer into our system during both training and test stages. Experimental results show that our approach produces better results than the state-of-the-art DL and non-DL methods on various synthetic and real datasets both visually and numerically.