CVPR 2026 Papers

Vision Transformers (ViTs) often need to be compressed for deployment on resource-constrained edge devices like drones and smart vehicles. However, existing model compression methods ignore that many edge devices only require the knowledge of specific classes for their applications. As a result, the derived all-class ViTs retain redundant knowledge and perform suboptimally on these classes. We discovered that simply replacing the calibration dataset with class-specific data does not suffice to address this issue, as these methods face two fundamental limitations. First, they overlook the existence of class-detrimental weights, which interfere with specialization, while removing them can improve class-specific performance. Second, the diversity of target classes and resource constraints on edge devices demand numerous customized models. Existing methods are time-consuming and computationally expensive, thus unscalable. In this work, we present NuWa, a cost-efficient method that addresses these challenges by deriving small ViTs from base ViTs for edge devices with specific class requirements. NuWa performs self-knowledge purification to prune class-detrimental weights and efficiently derives compact ViTs through closed-form optimization. Without post-pruning retraining, the derived edge ViTs surpass the base ViT in class-specific accuracy and accelerate inference. Comprehensive experiments demonstrate that NuWa outperforms state-of-the-art training-free pruning methods on class-specific tasks by up to 29.00% in accuracy. Compared with the best-performing training-dependent pruning method, NuWa achieves a 33.69x pruning speedup and reduces pruning cost by up to 99.83%, with only a 0.61% average accuracy loss.

In this work, we revisit several key design choices of modern Transformer-based approaches for feed-forward 3D Gaussian Splatting (3DGS) prediction. We argue that the common practice of regressing Gaussian means as depths along camera rays is suboptimal, and instead propose to directly regress 3D mean coordinates using only a self-supervised rendering loss.This formulation allows us to move from the standard encoder-only design to an encoder-decoder architecture with learnable Gaussian tokens, thereby _unbinding_ the number of predicted primitives from input image resolution and number of views. Our resulting method, __TokenGS__, demonstrates improved robustness to pose noise and multiview inconsistencies, while naturally supporting efficient test-time optimization in token space without degrading learned priors. TokenGS achieves state-of-the-art feed-forward reconstruction performance on both static and dynamic scenes, producing more regularized geometry and more balanced 3DGS distribution, while seamlessly recovering emergent scene attributes such as static-dynamic decomposition and scene flow.

Recent progress in 3D reconstruction has made it easy to create realistic digital twins from everyday environments. However, current digital twins remain largely static--limited to navigation and view synthesis without embodied interactivity. To bridge this gap, we introduce Dexterous World Model (DWM), an scene-action-conditioned video diffusion model enabling embodied interaction within static 3D scenes. Given a static 3D scene rendering and an egocentric hand motion sequence, DWM generates temporally coherent videos depicting plausible human-scene interactions. Our approach conditions video generation on (1) static scene renderings following a specified camera trajectory to ensure spatial consistency, and (2) egocentric hand mesh renderings that encode both geometry and motion cues in the egocentric view to model action-conditioned dynamics directly. We train our model on a synthetic human-scene interaction dataset and real-world object manipulation dataset, then evaluate it across both synthetic and real-world egocentric benchmarks. Experiments demonstrate that DWM enables realistic, physically grounded interactions, such as grasping, opening, or moving objects, while maintaining camera and scene consistency. This framework establishes the first step toward video diffusion-based interactive digital twins, enabling embodied simulation and 3D scene interactivity from egocentric actions.

Scene-level point cloud understanding remains challenging due to diverse geometries, imbalanced category distributions, and highly varied spatial layouts. Existing methods improve object-level performance but rely on static network parameters during inference, limiting their adaptability to dynamic scene data. We propose PointTPA, a Test-time Parameter Adaptation framework that generates input-aware network parameters for scene-level point clouds. PointTPA adopts a Serialization-based Neighborhood Grouping (SNG) to form locally coherent patches and a Dynamic Parameter Projector (DPP) to produce patch-wise adaptive weights, enabling the backbone to adjust its behavior according to scene-specific variations while maintaining a low parameter overhead. Integrated into the PTv3 structure, PointTPA demonstrates strong parameter efficiency by introducing two lightweight modules of less than 2% of the backbone's parameters. Despite this minimal parameter overhead, PointTPA achieves 78.4% mIoU on ScanNet validation, surpassing existing parameter-efficient fine-tuning (PEFT) methods across multiple benchmarks, highlighting the efficacy of our test-time dynamic network parameter adaptation mechanism in enhancing 3D scene understanding. The code is available at https://github.com/H-EmbodVis/PointTPA.

Reconstructing object geometry from radio frequency (RF) signals is fundamentally challenging due to the lensless imaging nature of RF sensing, which leads to low spatial resolution and high noise. Unlike light signals, RF signals can penetrate occlusions and thus capture information about hidden scenes. Existing Non-Line-of-Sight (NLoS) 3D neural reconstruction methods can recover coarse surfaces inside enclosed environments but often suffer from unstable optimization, noisy surface geometry, and surface ambiguity, failing to produce accurate zero-level sets from the signed distance field (SDF). These limitations largely stem from neglecting the role of Line-of-Sight (LoS) geometry outside the enclosed region, which provides valuable physical constraints for modeling signal propagation. In this paper, we introduce a Unified LoS and NLoS neural geometry reconstruction framework GeRaF 2.0 that leverages the outside LoS geometry to model and guide RF propagation from the LoS region into the NLoS region. By integrating visual LoS priors into the neural field formulation, GeRaF 2.0 achieves stable training and physically consistent reconstruction of both visible and hidden geometry, setting a new state-of-the-art in RF-based geometry reconstruction.

Since Transformers are introduced into vision architectures, their quadratic complexity has always been a significant issue that many research efforts aim to address. A representative approach involves grouping tokens, performing self-attention calculations within each group. To this end, various carefully designed grouping strategies have been proposed to enhance the performance of Vision Transformers. Here, we pose the following questions: Are these carefully designed grouping methods truly necessary? Is there a simpler and more unified token grouping method that can replace these diverse methods? Therefore, we propose the random grouping strategy, which involves a simple and fast random grouping strategy for vision tokens. We validate this approach on multiple baselines, and experiments show that random grouping almost outperforms all other grouping methods. When transferred to downstream tasks, such as object detection, random grouping demonstrates even more pronounced advantages. In response to this phenomenon, we conduct a detailed analysis of the advantages of random grouping from multiple perspectives and identify several crucial elements for the design of grouping strategies: positional information, head feature diversity, global receptive field, and fixed grouping pattern. We demonstrate that as long as these four conditions are met, vision tokens require only an extremely simple grouping strategy to efficiently and effectively handle various visual tasks. We also validate the effectiveness of our proposed random method across multiple modalities, including visual tasks, point cloud processing, and vision-language models. Code will be available at https://github.com/qhfan/random.

We present an approach for high-quality dynamic Gaussian Splatting from monocular videos. To this end, we in this work go one step further beyond previous methods to explicitly model continuous position and orientation deformation of dynamic Gaussians, using an SE(3) B-spline motion bases with a compact set of control points. To improve computational efficiency while enhancing the ability to model complex motions, an adaptive control mechanism is devised to dynamically adjust the number of motion bases and control points. Besides, we develop a soft segment reconstruction strategy to mitigate long-interval motion interference, and employ a multi-view diffusion model to provide multi-view cues for avoiding overfitting to training views. Extensive experiments demonstrate that our method outperforms state-of-the-art methods in novel view synthesis. Our code is available at https://github.com/hhhddddddd/se3bsplinegs.

Estimating 3D attributes directly from images has advanced rapidly with the Visual Geometry Grounded Transformer (VGGT), which predicts camera parameters, depth maps, and point clouds in a single forward pass. However, its 1.2B-parameter scale severely limits deployment on resource-constrained platforms such as UAVs and mobile AR devices. To address this limitation, we introduce QVGGT, a tailored quantization framework designed to compress VGGT. Our approach starts from the observation that transformer blocks within VGGT exhibit heterogeneous sensitivity to quantization. We thus analyze per-block quantization sensitivity and propose a selective mixed-precision strategy that allocates higher precision to the most fragile transformer blocks. To address the amplification of quantization error caused by high-variance camera and register tokens, we further introduce token filtering with camera information compensation, which removes these outliers from activation calibration and restores their geometric cues using a PCA-derived global compensation token. Finally, we develop a task-aware scale search mechanism that evaluates candidate quantization scales not only through layer reconstruction but also through multi-head supervision and cross-head geometric consistency among camera poses, depth maps, and point maps.Extensive experiments on multiple geometry perception benchmarks demonstrate that QVGGT achieves near-lossless W4A16 quantization, preserving the accuracy of all 3D prediction heads while delivering 3 4.9xmemory reduction and up to 2.8x real hardware speedup over FP32.Our approach makes high-fidelity 3D perception feasible on edge devices, enabling practical deployment of feed-forward 3D reconstruction models in real-world constrained environments.

Despite significant progress in Vision-Language Navigation (VLN), existing approaches still rely on dense RGB videos that produce excessive patch tokens and lack explicit spatial structure, resulting in substantial computational overhead and limited spatial reasoning. To address these issues, we introduce the Geometry-Aware BEV (GA-BEV) --a compact, 3D-grounded feature representation that integrates both explicit and implicit geometric cues into multimodal large language model (MLLM)-based navigation systems. We construct BEV spatial maps from RGB-D inputs by projecting visual features into 3D space and aggregating them into an agent-centric layout that preserves geometric consistency while reducing token redundancy. To further enrich geometric understanding, we incorporate features from a pretrained 3D foundation model into the BEV space, injecting structural priors learned from large-scale 3D reconstruction tasks. Together, these complementary cues--explicit depth-based projection and implicit learned priors--yield compact yet spatially expressive representations that substantially improve navigation efficiency and performance. Experiments show that our method achieves state-of-the-art results using only navigation data, without DAgger augmentation or mixed VQA training, demonstrating the robustness and data efficiency of the proposed GA-VLN framework.

Online 3D reconstruction from streaming inputs requires both long-term temporal consistency and efficient memory usage. Although causal variants of VGGT address this challenge through a key-value (KV) cache mechanism, the cache grows linearly with the stream length, creating a major memory bottleneck. Under limited memory budgets, early cache eviction significantly degrades reconstruction quality and temporal consistency. In this work, we observe that attention in causal transformers for 3D reconstruction exhibits intrinsic spatio-temporal sparsity. Based on this insight, we propose STAC, a Spatio-Temporally Aware Cache Compression framework for streaming 3D reconstruction with large causal transformers. STAC consists of three key components: (1) a Working Temporal Token Caching mechanism that preserves long-term informative tokens using decayed cumulative attention scores; (2) a Long-term Spatial Token Caching scheme that compresses spatially redundant tokens into voxel-aligned representations for memory-efficient storage; and (3) a Chunk-based Multi-frame Optimization strategy that jointly processes consecutive frames to improve temporal coherence and GPU efficiency. Extensive experiments show that STAC achieves state-of-the-art reconstruction quality while reducing memory consumption by nearly 10x and accelerating inference by 4x, substantially improving the scalability of real-time 3D reconstruction in streaming settings.

General 3D foundation models have started to lead the trend of unifying diverse vision tasks, yet most assume RGB-only inputs and ignore readily available geometric cues (e.g., camera intrinsics, poses, and depth maps). To address this issue, we introduce OmniVGGT, a novel framework that can effectively benefit from an arbitrary number of auxiliary geometric modalities during both training and inference. In our framework, a GeoAdapter is proposed to encode depth and camera intrinsics/extrinsics into a spatial foundation model. It employs zero-initialized convolutions to progressively inject geometric information without disrupting the foundation model's representation space. This design ensures stable optimization with negligible overhead, maintaining inference speed comparable to VGGT even with multiple additional inputs. Additionally, a stochastic multimodal fusion regimen is proposed, which randomly samples modality subsets per instance during training. This enables an arbitrary number of modality inputs during testing and promotes learning robust spatial representations instead of overfitting to auxiliary cues. Extensive experiments on monocular/multi-view depth estimation, multi-view stereo, and camera pose estimation demonstrate that OmniVGGT outperforms prior methods with auxiliary inputs and achieves state-of-the-art results even with RGB-only input. To further highlight its practical utility, we integrated OmniVGGT into vision-language-action (VLA) models. The enhanced VLA model by OmniVGGT not only outperforms the vanilla point-cloud-based baseline on mainstream benchmarks, but also effectively leverages accessible auxiliary inputs to achieve consistent gains on robotic tasks. Project Page: https://livioni.github.io/OmniVGGT-official/

We introduce PixARMesh, a method to autoregressively reconstruct complete 3D indoor scene meshes directly from a single RGB image. Unlike prior methods that rely on implicit signed distance fields and post-hoc layout optimization, PixARMesh jointly predicts object layout and geometry within a unified model, producing coherent and artist-ready meshes in a single forward pass. Building on recent advances in mesh generative models, we augment a point-cloud encoder with pixel-aligned image features and global scene context via cross-attention, enabling accurate spatial reasoning from a single image. Scenes are generated autoregressively from a unified token stream containing context, pose, and mesh, yielding compact meshes with high-fidelity geometry. Experiments on synthetic and real-world datasets show that PixARMesh achieves state-of-the-art reconstruction quality while producing lightweight, high-quality meshes ready for downstream applications.

Despite advances in physics-based 3D motion synthesis, current methods face key limitations: reliance on pre-reconstructed 3D Gaussian Splatting (3DGS) built from dense multi-view images with time-consuming per-scene optimization; physics integration via either inflexible, hand-specified attributes or unstable, optimization-heavy guidance from video models using Score Distillation Sampling (SDS); and naive concatenation of prebuilt 3DGS with physics modules, which ignores physical information embedded in appearance and yields suboptimal performance. To address these issues, we propose PhysGM, a feed-forward framework that jointly predicts 3D Gaussian representation and physical properties from a single image, enabling immediate simulation and high-fidelity 4D rendering. Unlike slow appearance-agnostic optimization methods, we first pre-train a physics-aware reconstruction model that directly infers both Gaussian and physical parameters. We further refine the model with Direct Preference Optimization (DPO), aligning simulations with the physically plausible reference videos and avoiding the high-cost SDS optimization. To address the absence of a supporting dataset for this task, we propose PhysAssets, a dataset of 50K+ 3D assets annotated with physical properties and corresponding reference videos. Experiments show that PhysGM produces high-fidelity 4D simulations from a single image in one minute, achieving a significant speedup over prior work while delivering realistic renderings.

Users frequently edit camera images post-capture to achieve their preferred photofinishing style. While editing in the RAW domain provides greater accuracy and flexibility, most edits are performed on the camera's display-referred output (e.g., 8-bit sRGB JPEG) since RAW images are rarely stored. Existing RAW reconstruction methods can recover RAW data from sRGB images, but these approaches are typically optimized for pixel-wise RAW reconstruction fidelity and tend to degrade under diverse rendering styles and editing operations. We introduce a plug-and-play, edit-aware loss function that can be integrated into any existing RAW reconstruction framework to make the recovered RAWs more robust to different rendering styles and edits. Our loss formulation incorporates a modular, differentiable image signal processor (ISP) that simulates realistic photofinishing pipelines with tunable parameters. During training, parameters for each ISP module are randomly sampled from carefully designed distributions that model practical variations in real camera processing. The loss is then computed in sRGB space between ground-truth and reconstructed RAWs rendered through this differentiable ISP. Incorporating our loss improves sRGB reconstruction quality by up to 1.5-2 dB PSNR across various editing conditions. Moreover, when applied to metadata-assisted RAW reconstruction methods, our approach enables fine-tuning for target edits, yielding further gains. Since photographic editing is the primary motivation for RAW reconstruction in consumer imaging, our simple yet effective loss function provides a general mechanism for enhancing edit fidelity and rendering flexibility across existing methods.

To ensure the robustness of 3D point cloud Deep Neural Network(3D DNN), 3D adversarial attack targeting the inference stage and backdoor attack targeting the training stage are well studied. The success of both attacks usually requires a specified permissions that attacker must have. However, the obtainable permissions are uncertain due to the deployment environment changes in practical scenarios. This renders existing separately designed adversarial attack or backdoor attack ineffective. To solve this issue, this paper proposes a unified attack that can adapt to both 3D point cloud backdoor attack and adversarial attack, named UAtt3D. Furthermore, by observing existing attacks, their way to promise attack stealthiness is to limit the undesirable perturbation. This strategy requires moving the point position as little as possible, which restricts the attack intensity and is not suitable for our unified attack. Meanwhile, this strategy will inevitably cause a quality decrease on 3D point cloud due to the remaining malicious perturbation. Therefore, our UAtt3D explores a new avenue to guarantee attack stealthiness which improves the quality of attacked 3D point cloud rather than decreasing it. In detail, to simultaneously consider feature movement of adversarial attack and backdoor feature learning of backdoor attack, a flexible isotropic resampling is designed. It realigns the position of most points based on surface approximation and rays sampling. By fine tuning the resampled point cloud, adversarial point cloud and backdoored point cloud are obtained. Several experiments suggest that the proposed UAtt3D achieves outstanding stealthiness comparing with existing adversarial attacks and backdoor attacks from the subjective and objective perspective. Meanwhile, its attack efficiency is competitive.

Recent 3D Gaussian Splatting (3DGS) dropout methods address overfitting under sparse-view conditions by randomly nullifying Gaussian opacities. However, we identify a neighbor compensation effect in these approaches: dropped Gaussians are often compensated by their neighbors, weakening the intended regularization. Moreover, these methods overlook the contribution of high-degree spherical harmonic coefficients (SH) to overfitting. To address these issues, we propose DropAnSH-GS, a novel anchor-based dropout strategy. Rather than dropping Gaussians independently, our method randomly selects certain Gaussians as anchors and simultaneously removes their spatial neighbors. This effectively disrupts local redundancies and encourages the model to learn more robust, globally informed representations. Furthermore, we extend the dropout to color attributes by randomly dropping higher-degree SH coefficients to concentrate appearance information in lower-degree SH. This strategy further mitigates overfitting and enables flexible post-training model compression via SH truncation. Experimental results demonstrate that DropAnSH-GS substantially outperforms existing dropout methods with negligible computational overhead and can be readily integrated into various 3DGS variants to enhance their performances.

Recent advancements in neural visual geometry, including transformer-based models such as VGGT and Pi3, have achieved impressive accuracy on 3D reconstruction tasks. However, their reliance on full attention makes them fundamentally limited by GPU memory capacity, preventing them from scaling to large, unordered image collections. We introduce MERG3R, a training-free divide-and-conquer framework that enables geometric foundation models to operate far beyond their native memory limits. MERG3R first reorders and partitions unordered images into overlapping, geometrically diverse subsets that can be reconstructed independently. It then merges the resulting local reconstructions through an efficient global alignment and confidence-weighted bundle adjustment procedure, producing a globally consistent 3D model. Our framework is model-agnostic and can be paired with existing neural geometry models. Across large-scale datasets--including 7-Scenes, NRGBD, Tanks & Temples, and Cambridge Landmarks--MERG3R consistently improves reconstruction accuracy, memory efficiency, and scalability, enabling high-quality reconstruction when the dataset exceeds memory capacity limits.

In high-stakes domains, small task-specific vision models are crucial due to their low computational requirements and the availability of numerous methods to explain their results. However, these explanations often reveal that the models do not align well with human domain knowledge, relying instead on spurious correlations. This might result in brittle behavior once deployed in the real-world. To address this issue, we introduce a novel and efficient method for aligning small task-specific vision models with human domain knowledge by leveraging the generalization capabilities of a Large Vision Language Model (LVLM). Our LVLM-Aided Visual Alignment (LVLM-VA) method provides a bidirectional interface that translates model behavior into natural language and maps human class-level specifications to image-level critiques, enabling effective interaction between domain experts and the model. Our method demonstrates substantial improvement in aligning model behavior with human specifications, as validated on both synthetic and real-world datasets. We show that it effectively reduces the model's dependence on spurious features and on group-specific biases, without requiring fine-grained feedback.

Generating dynamic and interactive 3D trees has wide applications in virtual reality, games, and world simulation. However, existing methods still face various challenges in generating structurally consistent and realistic 4D motion for complex real trees. In this paper, we propose DynamicTree, the first framework that can generate long-term, interactive 3D motion for 3DGS reconstructions of real trees. Unlike prior optimization-based methods, our approach generates dynamics in a fast feed-forward manner. The key success of our approach is the use of a compact sparse voxel spectrum to represent the tree movement. Given a 3D tree from Gaussian Splatting reconstruction, our pipeline first generates mesh motion using the sparse voxel spectrum and then binds Gaussians to deform the mesh. Additionally, the proposed sparse voxel spectrum can also serve as a basis for fast modal analysis under external forces, allowing real-time interactive responses. To train our model, we also introduce 4DTree, the first large-scale synthetic 4D tree dataset containing about 8.5k animated tree meshes with semantic labels and 100-frame motion sequences. Extensive experiments demonstrate that our method achieves realistic and responsive tree animations, significantly outperforming existing approaches in both visual quality and computational efficiency.