CVPR 2026 Papers

While Multimodal Large Language Models (MLLMs) excel at vision-language tasks, the cost of their language-driven training on internal visual foundational competence remains unclear. In this paper, we conduct a detailed diagnostic analysis to unveil a pervasive issue: visual representation degradation in MLLMs. Specifically, we find that compared to the initial visual features, the visual representation in the middle layers of LLM exhibits both a degradation in global function and patch structure. We attribute this phenomenon to a visual sacrifice driven by the singular text-generation objective, where the model compromises its visual fidelity to optimize for answer generation. We argue that a robust MLLM requires both strong cross-modal reasoning and core visual competence, and propose Predictive Regularization (PRe) to force degraded intermediate features to predict initial visual features, thereby maintaining the inherent visual attributes of the MLLM's internal representations. Extensive experiments confirm that mitigating this visual degradation effectively boosts vision-language performance, underscoring the critical importance of fostering robust internal visual representations within MLLMs for comprehensive multimodal understanding.

Can warping tokens, rather than pixels, help multimodal large language models (MLLMs) understand how a scene appears from a nearby viewpoint? While MLLMs perform well on visual reasoning, they remain fragile to viewpoint changes, as pixel-wise warping is highly sensitive to small depth errors and often introduces geometric distortions. Drawing on theories of mental imagery that posit part-level structural representations as the basis for human perspective transformation, we examine whether image tokens in ViT-based MLLMs serve as an effective substrate for viewpoint changes. We compare forward and backward warping, finding that backward token warping, which defines a dense grid on the target view and retrieves a corresponding source-view token for each grid point, achieves greater stability and better preserves semantic coherence under viewpoint shifts. Experiments on our proposed ViewBench benchmark demonstrate that token-level warping enables MLLMs to reason reliably from nearby viewpoints, consistently outperforming all baselines including pixel-wise warping approaches, spatially fine-tuned MLLMs, and a generative warping method.

Deep learning-based watermarking has made remarkable progress in recent years. To achieve robustness against various distortions, current methods commonly adopt a training strategy where a \underline s ingle \underline r andom \underline d istortion (SRD) is chosen as the noise layer in each training batch. However, the SRD strategy treats distortions independently within each batch, neglecting the inherent relationships among different types of distortions and causing optimization conflicts across batches. To address this issue, we propose a novel training strategy that enhances robustness and generalization via \underline meta -learning with \underline f eature \underline c onsistency (Meta-FC). Specifically, we randomly sample multiple distortions from the noise pool to construct a meta-training task, while holding out one distortion as a simulated "unknown" distortion for the meta-testing phase.Through meta-learning, the model is encouraged to identify and utilize neurons that exhibit stable activations across different types of distortions, mitigating the optimization conflicts caused by the random sampling of diverse distortions in each batch.To further promote the transformation of stable activations into distortion-invariant representations, we introduce a feature consistency loss that constrains the decoded features of the same image subjected to different distortions to remain consistent.Extensive experiments demonstrate that, compared to the SRD training strategy, Meta-FC improves the robustness and generalization of various watermarking models by an average of 1.59%, 4.71%, and 2.38% under high-intensity, combined, and unknown distortions. Code has been made \href https://github.com/007-li/Meta-FC open-source .

Event keypoint detection has garnered significant attention due to its ability to establish spatial associations through matching, which is fundamental for various computer vision tasks. However, achieving robust event keypoint detection remains challenging due to the difficulty in balancing the exploitation of event data and compatibility with established algorithms. Moreover, the limited use of co-visible information often results in excessive keypoint detection in non-matching regions, leading to incorrect matches. To address these challenges, we propose a novel Co-visible Focused 3D-guided 2D Event Keypoint Detection Network (EV-CGNet), which mainly consists of a 3D-guided 2D feature prototype learning (G2PL) module and a co-visible region-focused detector and descriptor learning (CDDL) module. The proposed method enjoys several merits. First, the proposed G2PL module can leverage fine-grained spatio-temporal cues from event points to guide the learning of event frame feature prototypes. Second, the proposed CDDL module can focus keypoint detection on co-visible regions, ensuring accurate matches. Comprehensive experimental evaluations on six challenging benchmarks show that our method significantly outperforms state-of-the-art event keypoint detection methods.

Sparse autoencoders (SAEs) promise a unified approach for mechanistic interpretability, concept discovery, and model steering in LLMs and LVLMs. However, realizing this potential requires learned features to be both interpretable and steerable. To that end, we introduce two new computationally inexpensive interpretability and steerability metrics for a systematic analysis of LVLM SAEs. This uncovers two observations; (i) a majority of SAE neurons exhibit either low interpretability or low steerability or both, rendering them ineffective for downstream use; and (ii) user-desired concepts are often absent in the SAE, thus limiting their practical utility. To address these limitations, we propose Concept Bottleneck Sparse Autoencoders (CB-SAE)--a novel post-hoc framework that prunes low-utility neurons and augments the latent space with a lightweight concept bottleneck aligned to a user-defined concept set. The resulting CB-SAE improves interpretability by +32.1% and steerability by +14.5% across LVLMs and image generation tasks.

Monocular depth estimation is essential for applications such as robotics. The complementary characteristics of event and image modalities have inspired fusion-based methods for robust depth estimation. However, existing methods rely on convolutional or attention-based architectures, which either have limited capacity for long-range modeling or incur high computational cost, making them less suitable for depth estimation over long sequences. Moreover, effective image-event fusion remains challenging, since most methods directly fuse features without addressing the domain gap and representational differences between raw events and images, resulting in semantic bias and degraded performance. In this work, we propose AIMDepth, an Asymmetric Image-Event Mamba framework for monocular depth estimation, built on state space models for linear complexity and accurate prediction. To alleviate input-domain misalignment, we introduce a Spectral Cross-modal Prior Guidance (SCPG) for bidirectional prior injection at the input level. To reduce the imbalance between sparse events and dense images, we design an asymmetric modal-aware Encoder (AME) with separate encoding paths and feature-level alignment. We further develop a Modality-interactive Local Refinement (ModiLocal) to enable hierarchical interaction and fine-grained alignment. Experiments on public datasets show that AIMDepth achieves state-of-the-art performance in complex environments.

Prototype-based Personalized Federated Learning (ProtoPFL) enables efficient multi-domain adaptation by communicating compact class prototypes, but directly sharing them poses privacy risks. A common defense involves per-example l_2 clipping before prototype computation to bound sensitivity, followed by isotropic Gaussian noise to enforce Local Differential Privacy (LDP). However, Isotropic Gaussian Prototype Perturbation (IGPP) typically over-perturbs discriminative dimensions and struggles to balance the clipping threshold with representation fidelity. In this paper, we propose VPDR, a client-side privacy plug-in that seamlessly integrates into existing ProtoPFLs. Motivated by the observation that dimension-wise class variance reflects discriminability, we introduce Variance-adaptive Prototype Perturbation (VPP), which allocates less noise to discriminative subspaces, preserving semantic separability while ensuring privacy. We further develop Distillation-guided Clipping Regularization (DCR), which enables feature norms to adaptively concentrate near the predefined clipping threshold while maintaining prediction consistency. Theoretical analysis shows that our groupwise mechanism provides privacy guarantees no weaker than the isotropic baseline under the same privacy constraints. Extensive experiments on multi-domain benchmarks demonstrate that VPDR achieves a superior privacy-utility trade-off, outperforming IGPP in personalized federated fine-tuning without sacrificing robustness against realistic attacks.

Optical Character Recognition (OCR) is fundamental to Vision-Language Models (VLMs) and high-quality data generation for LLM training. Yet, despite progress in average OCR accuracy, state-of-the-art VLMs still struggle with detecting sample-level errors and lack effective unsupervised quality control.We introduce Consensus Entropy (CE), a training-free, model-agnostic metric that estimates output reliability by measuring inter-model agreement entropy. The core insight is that correct predictions converge in output space, while errors diverge. Based on CE, we develop CE-OCR, a lightweight multi-model framework that verifies outputs by ensemble agreement, selects the best outputs, and further improves efficiency through adaptive routing. Experiments demonstrate that CE is robust for quality verification, improving F1 scores by 42.1% over VLM-as-Judge. CE-OCR achieves consistent OCR gains, outperforming self-consistency and single-model baselines at the same cost. Notably, CE requires no training or supervision, enabling plug-and-play integration.

Due to the lack of ground truth, existing methods of active stereo matching generally employ fully self-supervised learning to produce precise depth estimates. Although they can achieve promising results, their performance still has a noticeable gap compared with supervised models. To fill this gap, we propose a novel framework that synthesizes proxy labels to enable supervised training of deep active stereo networks without requiring any ground-truth depth. To expand the training data and generate disparity proxy labels, we develop an active 2D Gaussian Splatting (2DGS)-based synthesis method that explicitly models the scene geometry and the projected active pattern. Furthermore, to balance the varying contributions of different supervisions during training, we design a hybrid supervision regularization strategy that dynamically adjusts the loss weights to achieve stable optimization. We also contribute a real-world dataset captured by a handheld RealSense camera, along with our active 2DGS model, which facilitates future research on active stereo matching. Extensive qualitative and quantitative experiments demonstrate that our method achieves state-of-the-art performance on active stereo matching task. The code and dataset will be publicly released.

This paper tackles the problem of physics-aware human motion synthesis in a dynamic scene. Unlike existing works which mainly tend to generate physically unrealistic motions due to limited contact modeling, typically restricted to hands, in this paper, we introduce a physics-aware human motion generation framework that explicitly models the full spectrum of human-related forces, including human-object, human-scene, and internal body dynamics. Our method imposes soft physical constraints to maintain force and torque balance, ensuring physically grounded motion synthesis. We further propose a novel continuous distance-based force model that generalizes contact modeling to arbitrary surfaces, capturing interactions not only with static environments but also with dynamic, moving objects. Extensive experiments show that our approach significantly improves physical plausibility and generalizes well to complex scenes, setting a new benchmark for physically consistent human motion generation.

Reconstructing a physiologically plausible cardiac video from a single image remains a fundamental challenge in generative modeling, owing to the complex and nonlinear periodic dynamics of echocardiography. Previous image-to-video (I2V) approaches primarily focus on temporal continuity, yet often struggle to capture the intrinsic periodicity of cardiac motion, leading to limited temporal coherence and semantic consistency. We present EchoVDiff, a novel phase-aware diffusion model that reconstructs a full cardiac cycle from any single frame. Instead of direct pixel synthesis, EchoVDiff integrates physiological priors into a diffusion paradigm, learning interpretable mappings between cardiac phase, anatomy, and motion. By jointly modeling temporal rhythm and spatial semantics within a disentangled latent space, it achieves controllable and physiologically consistent generation. Extensive experiments on EchoNet-Dynamic and EchoNet-Pediatric demonstrate that EchoVDiff consistently surpasses state-of-the-art methods in both fidelity and temporal coherence. Remarkably, it enables accurate reconstruction of complete cardiac cycles from arbitrary phases, marking the first demonstration of single-frame-driven echocardiographic video generation.

This paper introduces LaS-Comp, a zero-shot and category-agnostic approach that leverages the rich geometric priors of 3D foundation models to enable 3D shape completion across diverse types of partial observations. Our contributions are threefold: First, LaS-Comp harnesses these powerful generative priors for completion through a complementary two-stage design: (i) an explicit replacement stage that preserves the partial observation geometry to ensure faithful completion; and (ii) an implicit refinement stage ensures seamless boundaries between the observed and synthesized regions. Second, our framework is training-free and compatible with different 3D foundation models. Third, we introduce Omni-Comp, a comprehensive benchmark combining real-world and synthetic data with diverse and challenging partial patterns, enabling a more thorough and realistic evaluation. Both quantitative and qualitative experiments demonstrate that our approach outperforms previous state-of-the-art approaches. Code and data will be released upon acceptance.

This work presents ODGS-SLAM, an omnidirectional simultaneous localization and mapping (SLAM) system utilizing 3D Gaussian Splatting (3DGS) as the unified representation for tracking and mapping. Thus, it reconstructs scene geometry from panoramic image sequences (RGB or RGBD) via splats while also estimating the camera poses. Such a framework is important to understand the full surrounding, e.g., for augmented reality applications or autonomous systems. We extended existing 3DGS-SLAM methods to handle omnidirectional input by including closed-form gradients for mapping and camera pose estimation, utilizing an equirectangular projection model. To reduce memory footprint, a key frame removal procedure based on graph analysis is proposed, enabling the application to handle larger input sizes. For evaluation, we provide a dataset of controlled real-world and synthetic test scenes (indoor and outdoor), employing a custom developed virtual camera lens. An extensive evaluation shows that, for camera tracking, the proposed method achieves statistically significant lower ATE RMSE scores compared to a recent omnidirectional SLAM system, as well as other 3DGS-SLAM frameworks, while reaching a similar mapping performance.

Large Vision-Language Models (VLMs) are increasingly deployed across diverse applications, such as AI agents, yet remain vulnerable to targeted adversarial attacks. However, the practical robustness of such attacks often remains unclear with limited evaluation under defenses. Diffusion-based purification (DBP), a widely adopted black-box defense, effectively blocks current attacks by removing adversarial perturbations via generative diffusion. Prior DBP evasion attacks target white-box image classifiers and are ill-suited to VLMs, incurring high computational costs and gradient instability when adapted. In this paper, we present PureProof, a black-box targeted attack on VLMs resilient to DBP. It consists of three core components. Stochastic Reverse Alignment (SRA) guides adversarial optimization via single-step reverse prediction, avoiding costly full-trajectory backpropagation. Adaptive Re-noising Augmentation (ARA) mitigates diffusion stochasticity through timestep-adaptive re-noising. Self-Consistency Regularization (SCR) stabilizes optimization by promoting local temporal coherence. Extensive experiments on open-source and commercial VLMs show that PureProof consistently outperforms prior attacks against DBP and achieves strong noise resilience, revealing critical vulnerabilities in VLMs and offering broader safety implications for real-world VLM deployments, including emerging agentic settings.

Reconstructing high-fidelity 3D head geometry from images is critical for a wide range of applications, yet existing methods face fundamental limitations. Traditional photogrammetry achieves exceptional detail but requires extensive camera arrays (25--200+ views), substantial computation, and manual cleanup in challenging areas like facial hair. Recent alternatives present a fundamental trade-off: foundation models enable efficient single-image reconstruction but lack fine geometric detail, while optimization-based methods achieve higher fidelity but require dense views and expensive computation. We bridge this gap with a hybrid approach that combines the strengths of both paradigms.Our method introduces a multi-view surface normal prediction model that extends monocular foundation models with cross-view attention to produce geometrically consistent normals in a feed-forward pass. We then leverage these predictions as strong geometric priors within an inverse rendering optimization framework to recover high-frequency surface details. Our approach outperforms state-of-the-art single-image and multi-view methods, achieving high-fidelity reconstruction on par with dense-view photogrammetry while reducing camera requirements and computational cost.

Egocentric 3D human motion estimation is essential for AR/VR experiences, yet remains challenging due to limited body coverage from the egocentric viewpoint, frequent occlusions, and scarce labeled data. We present EgoPoseFormer v2, a method that addresses these challenges through two key contributions: (1) a transformer-based model for temporally consistent and spatially grounded body pose estimation, and (2) an auto-labeling system that enables the use of large unlabeled datasets for training.The proposed model is fully differentiable, introduces identity-conditioned queries, multi-view spatial refinement, causal temporal attention, and supports both keypoints and parametric body representations under a constant compute budget.The proposed auto-labeling system scales learning to tens of millions of unlabeled frames via uncertainty-aware semi-supervised training. The system follows a teacher-student schema to generate pseudo-labels and guide training with uncertainty distillation, enabling the model to generalize to different environments. In experiments on the EgoBody3M benchmark, EgoPoseFormer v2 outperforms two state-of-the-art methods by 12.2% and 19.4% in accuracy, and reduces temporal jitter by 22.2% and 51.7%, respectively. Furthermore, our auto-labeling system additionally improves the wrist MPJPE by 13.1%.

Google Earth imagery, combined with building footprint databases, offers an efficient way to construct localized building datasets. However, the lack of orthorectification in these images leads to spatial misalignments between annotations and their corresponding roof locations. Adopting such misaligned data directly for model training can severely degrade segmentation performance. To address the challenge, we propose an Object-based Multi-stage Alignment Framework (OMAF) that generates high-quality corrected labels with minimal manual intervention. OMAF first employs a prior-regularized self-alignment method to produce high-confidence, object-level offset pseudo-labels, which are then used to train an instance-level offset regression model for label refinement. Experimental results on the challenging datasets demonstrate that OMAF effectively corrects misalignments and consistently boosts the mIoU of all baseline models by up to 40.6%. This work provides a practical and cost-effective solution for large-scale dataset construction and domain adaptation. Our code and dataset are publicly available at https://github.com/dayunyan/OMAF-Building-Alignment.

Contrastive learning is widely used for generating multimodal data representations by aligning embeddings of different modalities of the same data samples. This alignment is achieved through a loss function that treats matched and unmatched modality pairs as positive and negative samples within a data batch. However, when extending contrastive learning to scenarios involving more than two modalities, existing approaches either rely solely on fully unmatched modalities as negative samples, or fail to distinguish between partially and fully unmatched modalities, thereby overlooking the fine-grained contrasting relationships. To address this limitation, we propose Easy2Hard, a novel framework that explicitly separates partially and fully unmatched modalities. To learn from negative samples at improved granularity, Easy2Hard further introduces a sigmoid weighting curriculum that smoothly transitions the learning process from easy (partially unmatched) to hard (fully unmatched) contrasts. Comprehensive evaluations on five multimodal datasets across diverse domains demonstrate the superiority of our approach.

Diffusion Transformer (DiT)-based models have significantly advanced image editing by encoding conditional images and integrating them into transformer layers. However, most edits involve modifying only small regions, while current methods uniformly process and denoise all tokens at every timestep, causing redundant computation and potentially degrading unchanged areas. This raises a fundamental question: Is it truly necessary to regenerate every region during editing? To address this, we propose SpotEdit, a training-free diffusion editing framework that selectively updates only the modified regions. SpotEdit comprises two key components: SpotSelector identifies stable regions via perceptual similarity and skips their computation by reusing conditional image features; SpotFusion adaptively blends these features with edited tokens through a dynamic fusion mechanism, preserving contextual coherence and editing quality. By reducing unnecessary computation and maintaining high fidelity in unmodified areas, SpotEdit achieves efficient and precise image editing.