CVPR 2025 Papers

Three-dimensional scene inpainting is crucial for applications from virtual reality to architectural visualization, yet existing methods struggle with view consistency and geometric accuracy in 360deg unbounded scenes. We present AuraFusion360, a novel reference-based method that enables high-quality object removal and hole filling in 3D scenes represented by Gaussian Splatting. Our approach introduces (1) depth-aware unseen mask generation for accurate occlusion identification, (2) Adaptive Guided Depth Diffusion, a zero-shot method for accurate initial point placement without requiring additional training, and (3) SDEdit-based detail enhancement for multi-view coherence. We also introduce 360-USID, the first comprehensive dataset for 360deg unbounded scene inpainting with ground truth. Extensive experiments demonstrate that AuraFusion360 significantly outperforms existing methods, achieving superior perceptual quality while maintaining geometric accuracy across dramatic viewpoint changes.

Emerging from the pairwise attention in conventional Transformers, there is a growing interest in sparse attention mechanisms that align more closely with localized, contextual learning in the biological brain. Existing studies such as the Coordination method employ iterative cross-attention mechanisms with a bottleneck to enable the sparse association of inputs. However, these methods are parameter inefficient and fail in more complex relational reasoning tasks. To this end, we propose Associative Transformer (AiT) to enhance the association among sparsely attended input tokens, improving parameter efficiency and performance in various vision tasks such as classification and relational reasoning. AiT leverages a learnable explicit memory comprising specialized priors that guide bottleneck attentions to facilitate the extraction of diverse localized tokens. Moreover, AiT employs an associative memory-based token reconstruction using a Hopfield energy function. The extensive empirical experiments demonstrate that AiT requires significantly fewer parameters and attention layers outperforming a broad range of sparse Transformer models. Additionally, AiT outperforms the SOTA sparse Transformer models including the Coordination method on the Sort-of-CLEVR dataset.

We tackle the challenge of concurrent reconstruction at the part level with the RGB appearance and estimation of motion parameters for building digital twins of articulated objects using the 3D Gaussian Splatting (3D-GS) method. With two distinct sets of multi-view imagery, each depicting an object in separate static articulation configurations, we reconstruct the articulated object in 3D Gaussian representations with both appearance and geometry information at the same time. Our approach decoupled multiple highly interdependent parameters through a multi-step optimization process, thereby achieving a stable optimization procedure and high-quality outcomes. We introduce ArticulatedGS, a self-supervised, comprehensive framework that autonomously learns to model shapes and appearances at the part level and synchronizes the optimization of motion parameters, all without reliance on 3D supervision, motion cues, or semantic labels. Our experimental results demonstrate that, among comparable methodologies, our approach has achieved optimal outcomes in terms of part segmentation accuracy, motion estimation accuracy, and visual quality.

While existing vision and multi-modal foundation models can handle multiple computer vision tasks, they often suffer from significant limitations, including huge demand for data and computational resources during training and inconsistent performance across vision tasks at deployment time. To address these challenges, we introduce Argus (The name comes from Argus Panoptes -- a hundred-eyed giant with "all-seeing" capability in Greek mythology), a compact and versatile vision foundation model designed to support a wide range of vision tasks through a unified multitask architecture. Argus employs a two-stage training strategy: (i) multitask pretraining over core vision tasks with a shared backbone that includes a lightweight adapter to inject task-specific inductive biases, and (ii) scalable and efficient adaptation to new tasks by fine-tuning only the task-specific decoders. Extensive evaluations demonstrate that Argus, despite its relatively compact and training-efficient design of merely 100M backbone parameters (only 13.6% of which are trained using 1.6M images), competes with and even surpasses much larger models. Compared to state-of-the-art foundation models, Argus not only covers a broader set of vision tasks but also matches or outperforms the models with similar sizes on 12 tasks. We expect that Argus will accelerate the real-world adoption of vision foundation models in resource-constrained scenarios.

We introduce ArcPro, a novel learning framework built on architectural programs to recover structured 3D abstractions from highly sparse and low-quality point clouds. Specifically, we design a domain-specific language (DSL) to hierarchically represent building structures as a program, which can be efficiently converted into a mesh. We bridge feedforward and inverse procedural modeling by using a feedforward process for training data synthesis, allowing the network to make reverse predictions. We train an encoder-decoder on the points-program pairs to establish a mapping from unstructured point clouds to architectural programs, where a 3D convolutional encoder extracts point cloud features and a transformer decoder autoregressively predicts the programs in a tokenized form. Inference by our method is highly efficient and produces plausible and faithful 3D abstractions. Comprehensive experiments demonstrate that ArcPro outperforms both traditional architectural proxy reconstruction and learning-based abstraction methods. We further explore its potential when working with multi-view image and natural language inputs.

Inspired by the effectiveness of 3D Gaussian Splatting (3DGS) in reconstructing detailed 3D scenes within multi-view setups and the emergence of large 2D human foundation models, we introduce Arc2Avatar, the first SDS-based method utilizing a human face foundation model as guidance with just a single image as input. To achieve that, we extend such a model for diverse-view human head generation by fine-tuning on synthetic data and modifying its conditioning. Our avatars maintain a dense correspondence with a human face mesh template, allowing blendshape-based expression generation. This is achieved through a modified 3DGS approach, connectivity regularizers, and a strategic initialization tailored for our task. Additionally, we propose an optional efficient SDS-based correction step to refine the blendshape expressions, enhancing realism and diversity. Experiments demonstrate that Arc2Avatar achieves state-of-the-art realism and identity preservation, effectively addressing color issues by allowing the use of very low guidance, enabled by our strong identity prior and initialization strategy, without compromising detail.

Current learning-based Structure-from-Motion (SfM) methods struggle with videos of dynamic scenes from wide-angle cameras. We present AnyMap, a differentiable SfM framework that jointly addresses image distortion and motion estimation. By learning a general implicit camera model without predefined parameters, AnyMap effectively handles lens distortion, estimating multi-view consistent 3D geometry, camera poses, and (un)projection functions. To resolve the ambiguity where motion estimation can compensate for undistortion errors and vice versa, we introduce a low-dimensional motion representation consisting of a set of learnable basis trajectories, interpolated to produce regularized motion estimates. Experimental results show that our method produces accurate camera poses, excels in camera calibration and image rectification, and enables high-quality novel view synthesis. Our low-dimensional motion representation effectively disentangles undistortion with motion estimation, outperforming existing methods.

Recent image-based 3D reconstruction methods have achieved excellent quality for indoor scenes using 3D convolutional neural networks. However, they rely on a high-resolution grid in order to achieve detailed output surfaces, which is quite costly in terms of compute time, and it results in large mesh sizes that are more expensive to store, transmit, and render. In this paper we propose a new solution to this problem, using adaptive sampling. By re-formulating the final layers of the network, we are able to analytically bound the local surface complexity, and set the local sample rate accordingly. Our method, AniGrad, achieves an order of magnitude reduction in both surface extraction latency and mesh size, while preserving mesh accuracy and detail.

Generating animatable human avatars from a single image is essential for various digital human modeling applications. Existing 3D reconstruction methods often struggle to capture fine details in animatable models, while generative approaches for controllable animation, though avoiding explicit 3D modeling, suffer from viewpoint inconsistencies in extreme poses and computational inefficiencies. In this paper, we address these challenges by leveraging the power of generative models to produce detailed multi-view canonical pose images, which help resolve ambiguities in animatable human reconstruction. We then propose a robust method for 3D reconstruction of inconsistent images, enabling real-time rendering during inference. Specifically, we adapt a transformer-based Text-to-Video model to generate multi-view canonical pose images and normal maps, pretraining on a large-scale monocular video dataset to improve generalization. To handle view inconsistencies, we recast the reconstruction problem as a 4D task and introduce an efficient 3D modeling approach using 4D Gaussian Splatting. Experiments demonstrate that our method achieves photorealistic, real-time animation of 3D human avatars from in-the-wild images, showcasing its effectiveness and generalization capability.

Existing conditional Denoising Diffusion Probabilistic Models (DDPMs) with a Noise-Conditional Framework (NCF) remain challenging for 3D scene understanding tasks, as the complex geometric details in scenes increase the difficulty of fitting the gradients of the data distribution (the scores) from semantic labels. This also results in longer training and inference time for DDPMs compared to non-DDPMs. From a different perspective, we delve deeply into the model paradigm dominated by the Conditional Network. In this paper, we propose an end-to-end robust semantic Segmentation Network based on a Conditional-Noise Framework (CNF) of DDPMs, named CDSegNet. Specifically, CDSegNet models the Noise Network (NN) as a learnable noise-feature generator. This enables the Conditional Network (CN) to understand 3D scene semantics under multi-level feature perturbations, enhancing the generalization in unseen scenes. Meanwhile, benefiting from the noise system of DDPMs, CDSegNet exhibits strong robustness for data noise and sparsity in experiments. Moreover, thanks to CNF, CDSegNet can generate the semantic labels in a single-step inference like non-DDPMs, due to avoiding directly fitting the scores from semantic labels in the dominant network of CDSegNet. On public indoor and outdoor benchmarks, CDSegNet significantly outperforms existing methods, achieving state-of-the-art performance.

Diffusion-based policies have shown impressive performance in robotic manipulation tasks while struggling with out-of-domain distributions. Recent efforts attempted to enhance generalization by improving the visual feature encoding for diffusion policy. However, their generalization is typically limited to the same category with similar appearances. Our key insight is that leveraging affordances--manipulation priors that define "where" and "how" an agent interacts with an object--can substantially enhance generalization to entirely unseen object instances and categories. We introduce the Diffusion Policy with transferable Affordance (AffordDP), designed for generalizable manipulation across novel categories. AffordDP models affordances through 3D contact points and post-contact trajectories, capturing the essential static and dynamic information for complex tasks. The transferable affordance from in-domain data to unseen objects is achieved by estimating a 6D transformation matrix using foundational vision models and point cloud registration techniques. More importantly, we incorporate affordance guidance during diffusion sampling that can refine action sequence generation. This guidance directs the generated action to gradually move towards the desired manipulation for unseen objects while keeping the generated action within the manifold of action space. Experimental results from both simulated and real-world environments demonstrate that AffordDP consistently outperforms previous diffusion-based methods, successfully generalizing to unseen instances and categories where others fail.

We explore the task of geometric reconstruction of images captured from a mixture of ground and aerial views. Current state-of-the-art learning-based approaches fail to handle the extreme viewpoint variation between aerial-ground image pairs. Our hypothesis is that the lack of high-quality, co-registered aerial-ground datasets for training is a key reason for this failure. Such data is difficult to assemble precisely because it is difficult to reconstruct in a scalable way. To overcome this challenge, we propose a scalable framework combining pseudo-synthetic renderings from 3D city-wide meshes (e.g., Google Earth) with real, ground-level crowd-sourced images (e.g., MegaDepth). The pseudo-synthetic data simulates a wide range of aerial viewpoints, while the real, crowd-sourced images help improve visual fidelity for ground-level images where mesh-based renderings lack sufficient detail, effectively bridging the domain gap between real images and pseudo-synthetic renderings. Using this hybrid dataset, we fine-tune several state-of-the-art algorithms and achieve significant improvements on real-world, zero-shot aerial-ground tasks. For example, we observe that baseline DUSt3R localizes fewer than 5% of aerial-ground pairs within 5 degrees of camera rotation error, while fine-tuning with our data raises accuracy to nearly 56%, addressing a major failure point in handling large viewpoint changes. Beyond camera estimation and scene reconstruction, our dataset also improves performance on downstream tasks like novel-view synthesis in challenging aerial-ground scenarios, demonstrating the practical value of our approach in real-world applications.

In this work, we introduce the Adventurer series models where we treat images as sequences of patch tokens and employ uni-directional language models to learn visual representations. This modeling paradigm allows us to process images in a recurrent formulation with linear complexity relative to the sequence length, which can effectively address the memory and computation explosion issues posed by high-resolution and fine-grained images. In detail, we introduce two simple designs that seamlessly integrate image inputs into the causal inference framework: a global pooling token placed at the beginning of the sequence and a flipping operation between every two layers. Extensive empirical studies highlight that compared with the existing plain architectures such as DeiT and Vim, Adventurer offers an optimal efficiency-accuracy trade-off. For example, our Adventurer-Base attains a competitive test accuracy of 84.3% on the standard ImageNet-1k benchmark with 216 images/s training throughput, which is 3.8x and 6.2x faster than Vim and DeiT to achieve the same result. As Adventurer offers great computation and memory efficiency and allows scaling with linear complexity, we hope this architecture can benefit future explorations in modeling long sequences for high-resolution or fine-grained images.

Generalizable Neural Radiance Fields (GNeRF) are recognized as one of the most promising techniques for novel view synthesis and 3D model generation in real-world applications. However, like other generative models in computer vision, ensuring their adversarial robustness against various threat models is essential for practical use. The pioneering work in this area, NeRFool, introduced a state-of-the-art attack that targets GNeRFs by manipulating source views before feature extraction, successfully disrupting the color and density results of the constructed views. Building on this foundation, we propose IL2-NeRF (Iterative L2 NeRF Attack), a novel adversarial attack method that explores a new threat model (in the L2 domain) for attacking GNeRFs. We evaluated IL2-NeRF against two standard GNeRF models across three benchmark datasets, demonstrating similar performance compared to NeRFool, based on the same evaluation metrics proposed by NeRFool. Our results establish IL2-NeRF as the first adversarial method for GNeRFs under the L2 norm. We establish a foundational L2 threat model for future research, enabling direct performance comparisons while introducing a smoother, image-wide perturbation approach in Adversarial 3D Reconstruction. Our code is available at: https://github.com/The-NRC-SCAR-Group/IL2-NeRF

Point cloud video understanding is becoming increasingly important in fields such as robotics, autonomous driving, and augmented reality, as they can accurately represent object motion and environmental changes. Despite the progress made in self-supervised learning methods for point cloud video understanding, the limited availability of 4D data and the high computational cost of training 4D-specific models remain significant obstacles. In this paper, we investigate the potential of transferring pre-trained static 3D point cloud models to the 4D domain, identifying the limitations of static models that capture only spatial information while neglecting temporal dynamics. To address this, we propose a novel Cross-frame Spatio-temporal Adaptation (CSA) strategy by introducing the Point Tube Adapter as the embedding layer and the Geometric Constraint Temporal Adapter (GCTA) to enforce temporal consistency across frames. This strategy extracts both short-term and long-term temporal dynamics, effectively integrating them with spatial features and enriching the model's understanding of temporal changes in point cloud videos. Extensive experiments on 3D action and gesture recognition tasks demonstrate that our method achieves state-of-the-art performance, establishing its effectiveness for point cloud video understanding.

We introduce ActiveGAMER, an active mapping system that utilizes 3D Gaussian Splatting (3DGS) to achieve high-quality, real-time scene mapping and exploration. Unlike traditional NeRF-based methods, which are computationally demanding and restrict active mapping performance, our approach leverages the efficient rendering capabilities of 3DGS, allowing effective and efficient exploration in complex environments.The core of our system is a rendering-based information gain module that dynamically identifies the most informative viewpoints for next-best-view planning, enhancing both geometric and photometric reconstruction accuracy. ActiveGAMER also integrates a carefully balanced framework, combining coarse-to-fine exploration, post-refinement, and a global-local keyframe selection strategy to maximize reconstruction completeness and fidelity.Our system autonomously explores and reconstructs environments with state-of-the-art geometric and photometric accuracy and completeness, significantly surpassing existing approaches in both aspects. Extensive evaluations on benchmark datasets such as Replica and MP3D highlight ActiveGAMER's effectiveness in active mapping tasks.

Conventional frame-based imaging for active stereo systems has encountered major challenges in fast-motion scenarios. However, how to design a novel paradigm for high-speed depth sensing still remains an open issue. In this paper, we propose a novel problem setting, namely active event-based stereo vision, which provides the first insight of integrating binocular event cameras and an infrared projector for high-speed depth sensing. Technically, we first build a stereo camera prototype system and present a real-world dataset with over 21.5k spatiotemporal synchronized labels at 15 Hz, while also creating a realistic synthetic dataset with stereo event streams and 23.8k synchronized labels at 20 Hz. Then, we propose ActiveEventNet, a lightweight yet effective active event-based stereo matching neural network that learns to generate high-quality dense disparity maps from stereo event streams with low latency. Experiments demonstrate that our ActiveEventNet outperforms state-of-the-art methods meanwhile significantly reducing computational complexity. Our solution offers superior depth sensing compared to conventional stereo cameras in high-speed scenes, while also achieving the inference speed of up to 150 FPS with our prototype. We believe that this novel paradigm will provide new insights into future depth sensing systems. Our project can be available at https://github.com/jianing-li/active_event_based_stereo.

Prevailing Multimodal Large Language Models (MLLMs) encode the input image(s) as vision tokens and feed them into the language backbone, similar to how Large Language Models (LLMs) process the text tokens. However, the number of vision tokens increases quadratically as the image resolutions, leading to huge computational costs.In this paper, we consider improving MLLM's efficiency from two scenarios, (I) Reducing computational cost without degrading the performance. (II) Improving the performance with given budgets. We start with our main finding that the ranking of each vision token sorted by attention scores is similar in each layer except the first layer. Based on it, we assume that the number of essential top vision tokens does not increase along layers. Accordingly, for Scenario I, we propose a greedy search algorithm (G-Search) to find the least number of vision tokens to keep at each layer from the shallow to the deep. Interestingly, G-Search is able to reach the optimal reduction strategy based on our assumption. For Scenario II, based on the reduction strategy from G-Search, we design a parametric sigmoid function (P-Sigmoid) to guide the reduction at each layer of the MLLM, whose parameters are optimized by Bayesian Optimization. Extensive experiments demonstrate that our approach can significantly accelerate those popular MLLMs, e.g. LLaVA, and InternVL2 models, by more than 2 xwithout performance drops. Our approach also far outperforms other token reduction methods when budgets are limited, achieving a better trade-off between efficiency and effectiveness.

Large Vision Language Models (LVLMs) have achieved significant success across multi-modal tasks. However, the computational cost of processing long visual tokens can be prohibitively expensive on resource-limited devices. Previous methods have identified redundancy in visual tokens within the Large Language Model (LLM) decoder layers and have mitigated this by pruning tokens using a pre-defined or fixed ratio, thereby reducing computational overhead. Nonetheless, we observe that the impact of pruning ratio varies across different LLM layers and instances (image-prompt pairs). Therefore, it is essential to develop a layer-wise and instance-wise vision token pruning strategy to balance computational cost and model performance effectively. We propose ATP-LLaVA, a novel approach that adaptively determines instance-specific token pruning ratios for each LLM layer. Specifically, we introduce an Adaptive Token Pruning (ATP) module, which computes the importance score and pruning threshold based on input instance adaptively. The ATP module can be seamlessly integrated between any two LLM layers with negligible computational overhead. Additionally, we develop a Spatial Augmented Pruning (SAP) strategy that prunes visual tokens with both token redundancy and spatial modeling perspectives. Our approach reduces the average token count by 75% while maintaining performance, with only a minimal 1.9% degradation across seven widely used benchmarks.