CVPR 2026 论文列表

Continuous image tokenizers enable efficient visual generation, and those based on variational frameworks can learn smooth, structured latent representations through KL regularization. Yet this often leads to posterior collapse when using fewer tokens, where the encoder fails to encode informative features into the compressed latent space. To address this, we introduce **MacTok**, a **M**asked **A**ugmenting 1D **C**ontinuous **Tok**enizer that leverages image masking and representation alignment to prevent collapse while learning compact and robust representations. MacTok applies both random masking to regularize latent learning and DINO-guided semantic masking to emphasize informative regions in images, forcing the model to encode robust semantics from incomplete visual evidence. Combined with global and local representation alignment, MacTok preserves rich discriminative information in a highly compressed 1D latent space, requiring only 64 or 128 tokens. On ImageNet, MacTok achieves a competitive gFID of 1.44 at 256x256 and a state-of-the-art 1.52 at 512x512 with SiT-XL, while reducing token usage by up to 64x. These results confirm that masking and semantic guidance together prevent posterior collapse and achieve efficient, high-fidelity tokenization.

The rise of vision foundation models (VFMs) calls for systematic evaluation. A common approach pairs VFMs with large language models (LLMs) as general-purpose heads, followed by evaluation on broad Visual Question Answering (VQA) benchmarks. However, this protocol has two key blind spots: (i) Instruction tuning data may not align with VQA test distributions, meaning a wrong prediction can stem from such data mismatch rather than VFMs' visual shortcomings; (ii) VQA benchmarks often require multiple visual abilities in a single question, making it difficult to determine whether errors arise from the lack of all required abilities or just one key ability. To address these gaps, we introduce AVA-Bench, the first benchmark that explicitly disentangles 14 Atomic Visual Abilities (AVAs), foundational skills such as localization, depth estimation, and spatial understanding, which collectively support complex visual reasoning tasks. By decoupling AVAs and matching training and test distributions within each, AVA-Bench pinpoints exactly where a VFM excels or falters. Applying AVA-Bench to leading VFMs thus reveals distinctive "ability fingerprints," turning VFM selection from educated guesswork into principled engineering. Notably, we find that a 0.5B LLM yields similar VFM rankings as a 7B LLM while cutting GPU hours by 8x, enabling more efficient evaluation. By offering a comprehensive and transparent benchmark, we hope AVA-Bench lays the foundation for the next generation of VFMs.

Text-guided dynamic 3D character generation has advanced rapidly, yet producing high-quality motion that faithfully reflects rich textual descriptions remains challenging. Existing methods tend to generate limited sub-actions or incoherent motion due to fixed-length temporal inputs and discrete frame-wise representations that fail to capture rich motion semantics. We address these limitations by representing motion with continuous differentiable B-spline curves, enabling more effective motion generation without modifying the capabilities of the underlying generative model. Specifically, our closed-form, Laplacian-regularized B-spline solver efficiently compresses variable-length motion sequences into compact representations with a fixed number of control points. Further, we introduce a normal-fusion strategy for input shape adherence along with correspondence-aware and local-rigidity losses for motion-restoration quality. To train our model, we collate BIMO, a new dataset containing diverse variable-length 3D motion sequences with rich, high-quality text annotations. Extensive evaluations show that our feed-forward framework BiMotion generates more expressive, higher-quality, and better prompt-aligned motions than existing state-of-the-art methods, while also achieving faster generation. Our project page is at: https://wangmiaowei.github.io/BiMotion.github.io/.

Diagrams are widely used in daily life. However, offline diagrams typically exist in the form of images, lacking structured data representation, which significantly limits their reusability and editability. Current research mainly focuses on supporting basic query tasks for online diagrams and does not meet the semantic understanding and interaction requirements for complex offline diagrams. Although large language models (LLMs) possess powerful reasoning and knowledge integration capabilities, their performance in processing offline diagrams is unsatisfactory due to the inability to accurately understand the structure and content of offline diagrams. To address these issues,we propose DiagramDiff, a framework consisting of a high-precision diagram reconstruction model and an instance-level diagram element recognition model. The framework converts offline diagrams into standardized data structures, enabling LLMs to transition from being unable to understand offline diagrams to becoming intelligent assistants capable of performing tasks such as semantic reasoning, logical validation, and efficient diagram editing. We have constructed a dataset containing diagrams and their corresponding question and answering(Q&A) and editing tasks. Experiments demonstrate that DiagramDiff achieves state-of-the-art performance in diagram reconstruction and recognition tasks, significantly enhancing LLMs' understanding and interaction capabilities with offline diagrams.

Recent advances in Multi-modal Large Language Models (MLLMs) have predominantly focused on enhancing visual \perception to improve \accuracy. However, a critical question remains unexplored: Do models know when they do not know? Through a probing experiment, we reveal a severe \confidence miscalibration problem in MLLMs. To address this, we propose Confidence-Driven Reinforcement Learning (CDRL), which uses original-noise image pairs and a novel confidence-based reward to enhance perceptual sensitivity and robustly calibrate the model's confidence. Beyond training benefits, calibrated confidence enables more effective test-time scaling as a free lunch. We further propose Confidence-Aware Test-Time Scaling (CA-TTS), which dynamically coordinates Self-Consistency, Self-Reflection, and Visual Self-Check modules guided by confidence signals. An Expert Model acts in multiple roles (e.g., Planner, Critic, Voter) to schedule these modules and provide external verification. Our integrated framework establishes new state-of-the-art results with consistent 8.8% gains across four benchmarks. More ablation studies demonstrate the effectiveness of each module and scaling superiority. Our code will be released after the acception.

Autofocus in dynamic environments remains challenging for conventional frame-based sensors, which often fail under fast motion, low light, or high dynamic range conditions. Event cameras, with microsecond temporal resolution and asynchronous brightness detection, offer a promising alternative. However, typical event-based autofocus methods assume that the sharpest focus corresponds to the maximum event rate. In this paper, we reveal a counterintuitive yet consistent phenomenon: the true focus actually corresponds to a local minimum in the event-rate curve. We theoretically derive this behavior from the physics of event generation and show that as defocus blur increases, the event rate first rises and then declines, forming a dual-peak-valley structure across focal distances. Based on this insight, we propose an Event Structural Valley-based Autofocus (ESVA) framework that identifies the valley between two dominant peaks as the true focal position. ESVA integrates structural smoothing, consistency filtering, and a dual-peak constraint to robustly recover the valley under noise and motion disturbances. Extensive experiments on multiple synthetic and real-world datasets demonstrate that ESVA delivers accurate and stable focus estimation, consistently outperforming existing event-only autofocus methods without requiring image reconstruction or supervision.

Creating machines capable of understanding the world in 3D is essential in assisting designers that build and edit 3D environments and robots navigating and interacting within a three-dimensional space. Inspired by advances in language and image modeling, we investigate the potential of autoregressive models for a new modality: structured 3D scenes. To this end, we propose a unified LLM framework that aligns language, images, and 3D scenes and provide a detailed "cookbook" outlining critical design choices for achieving optimal training and performance addressing key questions related to data representation, modality-specific objectives, and more. We show how to tokenize complex 3D objects to incorporate into our structured 3D scene modality. We evaluate performance across four core 3D tasks - rendering, recognition, instruction-following, and question-answering - and four 3D datasets, synthetic and real-world. We show our model's effectiveness on reconstructing complete 3D scenes consisting of complex objects from a single image and on real-world 3D object recognition tasks. Project webpage: https://glab-caltech.github.io/kyvo/

Prototype-based methods enhance interpretability in image recognition by establishing intermediate part-level prototypes to build interpretable classifiers, enabling transparent decision-making through localized evidence and reference to prototypical examples. However, most existing methods typically depend on unimodal visual supervision and constrain prototypes within the visual embedding space, which inherently restricts their alignment with language-conditioned semantic representations. In this paper, we present PRISM (Prototype-based Reasoning with Inter-modal Semantic Mining), a framework for interpretable image recognition that leverages natural language as an auxiliary modality to guide the learning of class-specific part prototypes. PRISM introduces an information-theoretic attribution mechanism to extract semantically relevant image regions conditioned on textual descriptions. By aligning attribution maps with prototype activation patterns, PRISM implicitly anchors visual part prototypes to semantically meaningful image regions without requiring explicit concept annotations. To promote better localization and reduce redundancy among prototypes, we introduce a spatial compactness constraint that encourages each prototype to attend to non-overlapping and spatially concentrated regions. Experiments on multiple fine-grained benchmarks demonstrate that PRISM not only improves classification performance but also provides faithful and semantically grounded visual explanations.

All classifiers, including state-of-the-art vision models, possess invariants, partially rooted in the geometry of their linear mappings. These invariants, which reside in the null-space of the classifier, induce equivalent sets of inputs that map to identical outputs. The semantic content of these invariants remains vague, as existing approaches struggle to provide human-interpretable information. To address this gap, we present Semantic Interpretation of the Null-space Geometry (SING), a method that constructs equivalent images, with respect to the network, and assigns semantic interpretations to the available variations. We use a mapping from network features to multi-modal vision language models. This allows us to obtain natural language descriptions and visual examples of the induced semantic shifts. SING can be applied to a single image, uncovering local invariants, or to sets of images, allowing a breadth of statistical analysis at the class and model levels. For example, our method reveals that ResNet50 leaks relevant semantic attributes to the null space, whereas DINO-ViT, a ViT pretrained with self-supervised DINO, is superior in maintaining class semantics across the invariant space. Code is available at https://tinyurl.com/github-SING.

Vision-Language-Action (VLA) models have emerged as a powerful paradigm for general robotic manipulation. However, existing approaches typically omit intermediate reasoning steps and directly regress actions, limiting reasoning interpretability and performance in long-horizon or compositional tasks. Although recent studies introduce Chain-of-Thought (CoT) reasoning into VLA models, their effectiveness remains suboptimal due to two key issues: (1) generating a full reasoning trajectory at every timestep introduces substantial redundancy, thereby hinders real-time deployment and (2) reasoning is performed independently, neglecting temporal consistency, which leads to planning conflicts.We propose TRM-VLA, a temporal-aware reasoning and memorization framework that integrates explicit temporal modeling into the VLA reasoning process. TRM-VLA consists of two core components: (1) Keyframe-Triggered Reasoning (KTR), which identifies task progress and performs hierarchical CoT reasoning only at key decision points to reduce redundant inference; and (2) Granularity-adaptable Context Memory (GCM), which dynamically stores and retrieves historical reasoning trajectories to maintain inter-frame coherence and global context. Built upon a dual-system architecture--combining a multimodal foundation model for slow reasoning (System 2) with a diffusion-based policy for fast execution (System 1)--TRM-VLA learns to plan and act efficiently in a unified manner. Extensive experiments on LIBERO-90, SIMPLER, and four real-world robotic tasks demonstrate that TRM-VLA achieves state-of-the-art performance while improving reasoning efficiency.

Recent advances in cross-view geo-localization (CVGL) methods have shown strong potential for supporting unmanned aerial vehicle (UAV) navigation in GNSS-denied environments. However, existing work predominantly focuses on matching UAV views to onboard map tiles, which introduces an inherent trade-off between accuracy and storage overhead, and overlooks the importance of the UAV's heading during navigation. Moreover, the substantial discrepancies and varying overlaps in cross-view scenarios have been insufficiently considered, limiting their generalization to real-world scenarios. In this paper, we present Bearing-UAV, a purely vision-driven cross-view navigation method that jointly predicts UAV absolute location and heading from neighboring features, enabling accurate, lightweight, and robust navigation in the wild. Our method leverages global and local structural features and explicitly encodes relative spatial relationships, making it robust to cross-view variations, misalignment, and feature-sparse conditions. We also present Bearing-UAV-90k, a multi-city benchmark for evaluating cross-view localization and navigation. Extensive experiments show encouraging results that Bearing-UAV yields lower localization errors than previous matching/retrieval paradigms across diverse terrains. Our code is publicly available at https://github.com/liukejia121/bearinguav.

We introduce a framework for learning latent representations of 4D objects which are descriptive, faithfully capturing object geometry and appearance; compressive, aiding in downstream efficiency; and accessible, requiring minimal input, i.e., an unstructured dynamic point cloud, to construct. Specifically, Velox trains an encoder to compress spatiotemporal color point clouds into a set of dynamic tokens. These tokens are supervised using two complementary decoders: a 4D surface decoder, which models the time-varying surface distribution capturing the geometry; and a Gaussian decoder, which maps the tokens to 3D Gaussians, helping learn appearance. To demonstrate the utility of our representation, we evaluate it across three downstream tasks: video-to-4D generation, 3D tracking, and cloth simulation via image-to-4D generation, and observe strong performances in all settings.

Multi-rater medical image segmentation captures the inherent ambiguity of clinical interpretation, where diagnostic boundaries vary across experts and imaging devices. Existing approaches often reduce this diversity to consensus labels or treat rater differences as noise, resulting in overconfident and poorly calibrated models. We propose a harmonized probabilistic framework that disentangles acquisition artifacts from genuine annotator variability through adaptive feature conditioning and frequency-domain personalization. A lightweight Harmonizer Network implicitly models scanner-specific artifacts and performs dynamic feature modulation to standardize latent representations, ensuring that uncertainty reflects anatomy rather than noise. To represent rater-specific styles, we introduce High-Frequency Prompt Modules that operate in the spectral domain to encode annotator-dependent boundary precision and textural sensitivity. These prompts adaptively modulate harmonized features to produce personalized yet anatomically consistent segmentations. Furthermore, a Generalized Energy Distance (GED)-based regularization aligns the generative distribution with empirical annotation variability, promoting diversity where experts disagree and consensus where they converge. Experiments on LIDC-IDRI and NPC-170 show SOTA aggregated and individualized segmentation, with notable GED reductions and improved Dice scores, especially on noisy cases. Beyond accuracy, the model exhibits clinically meaningful uncertainty, confidence rises in agreement regions and declines in ambiguous areas, supporting its use as a reliable and interpretable tool for multi-expert clinical workflows.

Large Vision-Language Models (LVLMs) usually suffer from prohibitive computational and memory costs due to the quadratic growth of visual tokens with image resolution. Existing token compression methods, while varied, often lack a high-level semantic understanding, leading to suboptimal merges, information redundancy, or context loss. To address these limitations, we introduce CORE (Compact Object-centric REpresentations), a new paradigm for visual token compression. CORE leverages an efficient segmentation decoder to generate object masks, which serve as a high-level semantic prior to guide the merging of visual tokens into a compact set of object-centric representations. Furthermore, a novel centroid-guided sorting mechanism restores a coherent spatial order to the merged tokens, preserving vital positional information. Extensive experiments show that CORE not only establishes a new state-of-the-art on six authoritative benchmarks for fixed-rate compression, but also achieves dramatic efficiency gains in adaptive-rate settings. Even under extreme compression, after aggressively retaining with only 2.2% of all visual tokens, CORE still maintains 97.4% of baseline performance. Our work demonstrates the superiority of object-centric representations for efficient and effective LVLM processing. The code is available at https://github.com/jingyulei/CORE.

Masked Image Modeling (MIM) has become a ubiquitous self-supervised vision paradigm. In this work, we show that MIM objectives cause the learned representations to retain non-semantic information, which ultimately hurts performance during inference. We introduce a model-agnostic score for semantic invariance using Principal Component Analysis (PCA) on real and synthetic non-semantic images. Based on this score, we propose a simple method, Semantically Orthogonal Artifact Projection (SOAP), to directly suppress non-semantic information in patch representations, leading to consistent improvements in zero-shot performance across various MIM-based models. SOAP is a post-hoc suppression method, requires zero training, and can be attached to any model as a single linear head.

Digital humans are fundamental to immersive interaction, yet creating a unified model for holistic modalities, including text, audio, motion, and visual content, remains an open challenge. In this paper, we present Archon, a fully pretrained, human-centric unified multimodal model for holistic avatar generation. Archon unifies seven modalities with modality-specific tokenizers, and a native autoregressive unified multimodal model pretrained on synchronized modalities and 72 diverse tasks to model holistic joint distributions. To address the token explosion challenge in high-fidelity talking videos, we introduce a memory-efficient semantic video reparameterization, achieving 4x token reduction while preserving fine-grained dynamics, coupled with a semantic-driven video diffusion decoder. We further propose a "Thinking in Modality" that decomposes ambiguous cross-modal tasks into stepwise thinking in an alternative chain of modality, progressively enhancing fidelity and controllability. Extensive experiments demonstrate that Archon achieves superior or comparable performance across diverse digital human generation tasks, validating the effectiveness of our unified framework.

Latent diffusion has become the default paradigm for visual generation, yet we observe a persistent reconstruction-generation trade-off as latent dimensionality increases: higher-capacity autoencoders improve reconstruction fidelity but generation quality eventually declines. We trace this gap to the different behaviors in high-frequency tokenization and detokenization. Through controlled perturbations in both RGB and latent domains, we analyze encoder/decoder behaviors and find that decoders depend strongly on high-frequency latent components to recover details, whereas encoders under-represent high-frequency contents, yielding insufficient exposure and underfitting in high-frequency bands for diffusion model training. To address this issue, we introduce FreqWarm, a plug-and-play frequency warm-up curriculum that increases early-stage exposure to high-frequency latent signals during diffusion or flow-matching training -- without modifying or retraining the autoencoder. Applied across several high-dimensional tokenizers, FreqWarm consistently improves generation quality: decreasing gFID by 14.11 on Wan2.2-VAE, 6.14 on LTX-VAE, and 4.42 on DC-AE-f32, while remaining architecture-agnostic and compatible with diverse backbones. Our study shows that explicitly managing frequency exposure can successfully turn high-dimensional latent spaces into more diffusible targets.

Embodied agents face a fundamental limitation: once deployed in real-world environments, they cannot easily acquire new knowledge to improve task performance. In this paper, we propose Dejavu, a general post-deployment learning framework that augments a frozen Vision-Language-Action (VLA) policy with retrieved execution memories through an Experience Feedback Network (EFN). EFN identifies contextually relevant prior action experiences and conditions action prediction on the retrieved guidance. We train EFN with reinforcement learning and semantic similarity rewards, encouraging the predicted actions to align with past behaviors under the current observation. During deployment, EFN continually expands its memory with new trajectories, enabling the agent to exhibit "learning from experience." Experiments across diverse embodied tasks show that EFN improves adaptability, robustness, and success rates over frozen baselines. Our Project Page is https://dejavu2025.github.io/.

Understanding objects in dynamic 4D environments via natural language is crucial yet underexplored. While existing methods focus on static 3D referring segmentation or open-vocabulary 4D querying, they struggle to ground complex spatio-temporal referring expressions in explicit 4D reconstructions. We introduce Spatio-Temporal Referring Segmentation in 4D Gaussian Splatting (STRS-4DGS), a novel task aiming to jointly identify and segment a target instance across space and time given a referring expression. To tackle this, we propose ST4R-Splat, the first framework for STRS-4DGS. Specifically, our framework incorporates an Instance-Aware 4D Gaussian Referring Field that assigns time-invariant embeddings for robust spatial grounding, and an Instance-Level Temporal State Mapping module that enables view-independent temporal localization directly in feature space. To provide rich spatio-temporal semantic supervision, we develop an automatic, MLLM-based captioning pipeline that generates decoupled spatial and temporal textual descriptions. Evaluated on our newly constructed STRS-4DGS benchmark, our method significantly outperforms adapted state-of-the-art baselines across both time-agnostic and time-sensitive metrics, establishing a strong foundation for language-driven 4D scene understanding.