CVPR 2026 论文列表

Multimodal large language models (MLLMs) have advanced from image-level reasoning to pixel-level grounding, but extending these capabilities to videos remains challenging as models must achieve spatial precision and temporally consistent reference tracking. Existing video MLLMs often rely on a static segmentation token ([SEG]) for frame-wise grounding, which provides semantics but lacks temporal context, causing spatial drift, identity switches, and unstable initialization when objects move or reappear. We introduce SPARROW, a pixel-grounded video MLLM that unifies spatial accuracy and temporal stability through two key components: (i) Target-Specific Tracked Features (TSF), which inject temporally aligned referent cues during training, and (ii) a dual-prompt design that decodes box ([BOX]) and segmentation ([SEG]) tokens to fuse geometric priors with semantic grounding. SPARROW is supported by a curated referential video dataset of 30,646 videos and 45,231 Q&A pairs and operates end-to-end without external detectors via a class-agnostic SAM2-based proposer. Integrated into three recent open-source video MLLMs (UniPixel, GLUS, and VideoGLaMM), SPARROW delivers consistent gains across six benchmarks, improving up to +8.9 J&F on RVOS, +5 mIoU on visual grounding, and +5.4 CLAIR on GCG. These results demonstrate that SPARROW substantially improves referential stability, spatial precision, and temporal coherence in pixel-grounded video understanding. Project page: https://risys-lab.github.io/SPARROW/

Ultra-High-Definition (UHD) image restoration is trapped in a scalability crisis: existing models, bound to pixel-wise operations, demand unsustainable computation. While state space models (SSMs) like Mamba promise linear complexity, their pixel-serial scanning remains a fundamental bottleneck for the millions of pixels in UHD content. We ask: must we process every pixel to understand the image? This paper introduces C^2SSM, a visual state space model that breaks this taboo by shifting from pixel-serial to cluster-serial scanning. Our core discovery is that the rich feature distribution of a UHD image can be distilled into a sparse set of semantic centroids via a neural-parameterized mixture model. C^2SSM leverages this to reformulate global modeling into a novel dual-path process: it scans and reasons over a handful of cluster centers, then diffuses the global context back to all pixels through a principled similarity distribution, all while a lightweight modulator preserves fine details. This cluster-centric paradigm achieves a decisive leap in efficiency, slashing computational costs while establishing new state-of-the-art results across five UHD restoration tasks. More than a solution, C^2SSM charts a new course for efficient large-scale vision: scan clusters, not pixels. The code is available at https://github.com/5chen/C2SSM.

Text-image to video generation aims to synthesize a video conditioned on the given text-image inputs. Nevertheless, existing methods generally assume that the semantic information carried in the input text and image tends to be perfectly paired and temporally aligned, occurring simultaneously in the generated video. As such, existing literature struggles with out-of-distribution (OOD) "unpaired" text-image inputs in the more universal and realistic scenario where i) the semantic information carried by the text and image may occur at different timestamps and ii) the condition image can appear at an arbitrary position rather than the first frame of the synthesized video. Video generation under this OOD setting poses an urgent need to conduct reasoning over the intrinsic connections between the given textual description and referred image, which is challenging and remains unexplored. To address the challenge, in this paper we study the problem of unpaired text-image to video generation for the first time, proposing ReasonDiff, a novel model for accurate video generation from unpaired text-image inputs. Specifically, ReasonDiff designs a VisionNarrator module to harness the powerful reasoning abilities of a multi-modal LLM to analyze the unpaired text-image inputs, producing coherent per-frame narratives that temporally align them. Building upon this VisionNarrator module, ReasonDiff further introduces a novel AlignFormer module, which employs a Multi-stage Temporal Anchor Attention mechanism to predict frame-wise latent representations. These reasoning-enhanced latents are subsequently fused with the condition frame, providing structured guidance throughout the video generation process. Extensive experiments and ablation studies demonstrate that ReasonDiff beats state-of-the-art baselines in terms of video generation quality with unpaired text-image inputs.

Understanding and answering questions based on a user's pointing gesture is essential for next-generation egocentric AI assistants. However, current Multimodal Large Language Models (MLLMs) struggle with such tasks due to the lack of gesture-rich data and their limited ability to infer fine-grained pointing intent from egocentric video.To address this, we introduce EgoPointVQA, a dataset and benchmark for gesture-grounded egocentric question answering, comprising 4000 synthetic and 400 real-world videos across multiple deictic reasoning tasks.Built upon it, we further propose Hand Intent Tokens (HINT), which encode tokens derived from 3D hand keypoints using an off-the-shelf reconstruction model and interleaves them with the model input to provide explicit spatial and temporal context for interpreting pointing intent.We show that our model outperforms others in different backbones and model sizes.In particular, HINT-14Bachieves 68.1% accuracy, on average over 6 tasks, surpassing the state-of-the-art, InternVL3-14B, by 6.6%.To further facilitate the open research, we will release the code, model, and dataset.

Monocular 3D object detection from a single RGB image remains challenging due to two fundamental challenges: the ill-posed nature of 3D localization, where multiple plausible configurations can correspond to the same 2D observation, and unreliable confidence estimation that fails to reflect true localization accuracy. Existing methods predict deterministic 3D boxes that often collapse to implausible mean estimates and rely on absolute confidence scores that are highly sensitive to localization errors. This paper introduces RARE, a unified framework that addresses both challenges through learning to rank and retrieve. RARE formulates confidence estimation as a ranking problem, learning to order detections by their relative quality rather than regressing absolute values. It provides more robust and stable confidence estimates that are less sensitive to localization uncertainty. Building on this improved confidence estimator, RARE learns to construct a query set for each object that predicts multiple diverse and plausible 3D configurations, and retrieves the top-ranked prediction. It explicitly models the multimodal nature of monocular 3D perception and produces more plausible localizations. Extensive experiments demonstrate the effectiveness of RARE. The code is available at https://github.com/HyeonjeongPark37/RARE.

Detecting visual anomalies in industrial inspection often requires training with only a few normal images per category. Recent few-shot methods achieve strong results employing foundation-model features, but typically rely on memory banks, auxiliary datasets, or multi-modal tuning of vision-language models. We therefore question whether such complexity is necessary given the feature representations of vision foundation models. To answer this question, we introduce SubspaceAD, a training-free method, that operates in two simple stages. First, patch-level features are extracted from a small set of normal images by a frozen DINOv2 backbone. Second, a Principal Component Analysis (PCA) model is fit to these features to estimate the low-dimensional subspace of normal variations. At inference, anomalies are detected via the reconstruction residual with respect to this subspace, producing interpretable and statistically grounded anomaly scores. Despite its simplicity, SubspaceAD achieves state-of-the-art performance across one-shot and few-shot settings without training, prompt tuning, or memory banks. In the one-shot anomaly detection setting, SubspaceAD achieves image-level and pixel-level AUROC of 98.0% and 97.6% on the MVTec-AD dataset, and 93.3% and 98.3% on the VisA dataset, respectively, surpassing prior state-of-the-art results.

Part-level pointing is important for fine-grained interaction and reasoning, yet existing Multimodal Large Language Models (MLLMs) remain limited to instance-level pointing. Part-level pointing presents unique challenges: annotation is costly, parts are long-tail distributed, and many are difficult to specify precisely in language. We introduce POinting at Parts (POP), a training-free, plug-and-play approach that addresses these challenges under a few-shot setup. POP fuses textual and visual attention maps with self-supervised visual correspondences from query image and few-shot examples. On average across the three evaluated datasets, POP achieves accuracy gains of up to 8.9 points in the one-shot setting and 16.4 points in the three-shot setting for the pointing-capable MLLMs--Qwen2.5-VL, Ovis2.5, and Molmo. Notably, even MLLMs without pointing capability benefit significantly from the proposed approach. These results establish a simple yet effective path toward fine-grained spatial grounding in MLLMs.

Vision-Language-Action (VLA) models have recently emerged, demonstrating strong generalization in robotic scene understanding and manipulation. However, when confronted with long-horizon tasks that require defined goal states, such as LEGO assembly or object rearrangement, existing VLA models still face challenges in coordinating long-horizon planning with precise manipulation.Therefore, we aim to endow a VLA model with the capability to infer the "how" process from the "what" outcomes, transforming goal states into executable procedures. In this paper, we introduce ManualVLA, a unified VLA framework built upon a Mixture-of-Transformers (MoT) architecture, enabling coherent collaboration between multimodal manual generation and action execution. Unlike prior VLA models that directly map sensory inputs to actions, we first equip ManualVLA with a planning expert that generates intermediate manuals consisting of images, visual prompts, and textual instructions. Building upon these multimodal manuals, we design a Manual Chain-of-Thought (ManualCoT) reasoning process that feeds them into the action expert, where each manual step provides explicit control conditions, while its latent representation offers implicit guidance for accurate manipulation. To alleviate the burden of data collection, we develop a high-fidelity digital-twin toolkit based on 3D Gaussian Splatting, which automatically generates manual data for planning expert training. ManualVLA demonstrates strong real-world performance, achieving an average success rate 32% higher than the previous hierarchical SOTA baseline on LEGO assembly and object rearrangement tasks.

Motion blur is a common degradation in dynamic imaging. Recent studies have moved beyond restoring a single sharp image from a blurred input and instead target blur decomposition: recovering a temporally continuous sharp video sequence from one motion-blurred image. Event cameras, with their microsecond temporal resolution, can effectively alleviate motion ambiguity. However, existing event-based methods often fail to explicitly model time-aligned event-image features. How to accurately exploit event data to reconstruct frames at different time instants remains largely underexplored. In this paper, we propose TSANet, an event-based blur-to-video decomposition method that time-specializes both event features and image features for alignment. Specifically, we introduce a Relative Time-Encoded Attention module that steers event features toward motion information relevant to a given target time, and a Timesurface Dynamic Warping module that warps image features into the spatial configuration corresponding to that time. With time-specialized motion and image features explicitly aligned at arbitrary query times, our framework can decompose a single blurred image into a high-frame-rate sharp video sequence. In addition, we collect a new dataset containing real events and high-quality color videos, and synthesize blurred inputs by averaging sharp frames to evaluate our method. Experiments on multiple datasets with both synthetic and real events demonstrate that our approach consistently outperforms previous state-of-the-art methods on the blur decomposition task.

As Vision-Language Models (VLMs) move into interactive, multi-turn use, safety concerns intensify for multimodal multi-turn dialogue, which is characterized by concealment of malicious intent, contextual risk accumulation, and cross-modal joint risk. These characteristics limit the effectiveness of content moderation approaches designed for single-turn or single-modality settings. To address these limitations, we first construct the Multimodal Multi-turn Dialogue Safety (MMDS) dataset, comprising 4,484 annotated dialogues and a comprehensive risk taxonomy with 8 primary and 60 subdimensions. As part of MMDS construction, we introduce Multimodal Multi-turn Red Teaming (MMRT), an automated framework for generating unsafe multimodal multi-turn dialogues. We further propose LLaVAShield, which audits the safety of both user inputs and assistant responses under specified policy dimensions in multimodal multi-turn dialogues. Extensive experiments show that LLaVAShield significantly outperforms state-of-the-art VLMs and existing content moderation tools while demonstrating strong generalization and flexible policy adaptation. Additionally, we analyze vulnerabilities of mainstream VLMs to harmful inputs and evaluate the contribution of key components, advancing understanding of safety mechanisms in multimodal multi-turn dialogues.

Nighttime photography is susceptible to flare caused by strong light sources, which degrades visual quality and disrupts structural information required by downstream vision tasks. Existing nighttime flare removal methods generally lack semantic priors for flare-occluded regions and thus tend to introduce artifacts and lose details under severe degradation. To address this problem, we propose a language-guided one-step diffusion framework that explicitly aligns flare-occluded regions with the underlying scene content at the semantic level. Specifically, we develop the first flare-specific vision-language model, Flare-VLM, which extracts fine-grained textual descriptions to guide one-step diffusion for high-quality restoration of severely damaged areas. Then, we propose semantics-aware distribution distillation to constrain the noise distribution with high-level semantics, suppressing redundant perturbations on clean backgrounds and improving the stability of distillation. In addition, we design an instruction-driven data synthesis pipeline to generate geometrically and semantically aligned nighttime flare samples, narrowing the gap to real degradations. Experimental results demonstrate that the proposed method achieves better restoration and enhances the performance of downstream vision tasks.

Large Vision-Language Models (LVLMs) have achieved remarkable progress in visual-textual understanding, yet their reliability is critically undermined by hallucinations, i.e., the generation of factually incorrect or inconsistent responses.While recent studies using steering vectors demonstrated promise in reducing hallucinations, a notable challenge remains: they inadvertently amplify the severity of residual hallucinations. We attribute this to their exclusive focus on the decoding stage, where errors accumulate autoregressively and progressively worsen subsequent hallucinatory outputs.To address this, we propose Prefill-Time Intervention (PTI), a novel steering paradigm that intervenes only once during the prefill stage, enhancing the initial Key-Value (KV) cache before error accumulation occurs.Specifically, PTI is modality-aware, deriving distinct directions for visual and textual representations. This intervention is decoupled to steer keys toward visually-grounded objects and values to filter background noise, correcting hallucination-prone representations at their source.Extensive experiments demonstrate PTI's significant performance in mitigating hallucinations and its generalizability across diverse decoding strategies, LVLMs, and benchmarks. Moreover, PTI is orthogonal to existing decoding-stage methods, enabling plug-and-play integration and further boosting performance.

Current training-free methods tackle MLLM hallucination with separate strategies: either enhancing visual signals or suppressing text inertia. However, these separate methods are insufficient due to critical trade-offs: simply enhancing vision often fails against strong language prior, while suppressing language can introduce extra image-irrelevant noise. Moreover, we find their naive combination is also ineffective, necessitating a unified framework. We propose such a framework by focusing on the core asset: the vision token. Our design leverages two key insights: (1) augmented images offer complementary visual semantics, and (2) removing vision tokens (information-gap) isolates hallucination tendencies more precisely than distorting images (modality-gap). Based on these, our framework uses vision tokens in two distinct ways, both operating on latent representations: our Synergistic Visual Calibration (SVC) module incorporates augmented tokens to strengthen visual representations, while our Causal Representation Calibration (CRC) module uses pruned tokens to create latent-space negative samples for correcting internal model biases. By harmonizing these two roles, our framework effectively restores the vision-language balance, significantly reducing object hallucinations, improving POPE accuracy by an average of 2% absolute on LLaVA-1.5 across multiple benchmarks with only a 1.06x inference latency overhead.

Recent advancements in video generation models have significantly improved their ability to follow text prompts. However, the customization of dynamic visual effects, defined as temporally evolving and appearance-driven visual phenomena like object crushing or explosion, remains underexplored. Prior works on motion customization or control mainly focus on low-level motions of the subject or camera, which can be guided using explicit control signals such as motion trajectories. In contrast, dynamic visual effects involve higher-level semantics that are more naturally suited for control via text prompts. However, it is hard and time-consuming for humans to craft a single prompt that accurately specifies these effects, as they require complex temporal reasoning and iterative refinement over time. To address this challenge, we propose P-Flow, a novel training-free framework for customizing dynamic visual effects in video generation without modifying the underlying model. By leveraging the semantic and temporal reasoning capabilities of vision-language models, P-Flow performs test-time prompt optimization, refining prompts based on the discrepancy between the visual effects of the reference video and the generated output. Through iterative refinement, the prompts evolve to better induce the desired dynamic effect in novel scenes. Experiments demonstrate that P-Flow achieves high-fidelity and diverse visual effect customization and outperforms other models on both text-to-video and image-to-video generation tasks. Code is available at https://github.com/showlab/P-Flow.

Vision-language models (VLMs) achieve strong multimodal performance, yet how computation is organized across populations of neurons remains poorly understood. In this work, we study VLMs through the lens of neural topology, representing each layer as a within-layer correlation graph derived from neuron-neuron co-activations. This view allows us to ask whether population-level structure is behaviorally meaningful, how it changes across modalities and depth, and whether it identifies causally influential internal components under intervention. We show that correlation topology carries recoverable behavioral signal; moreover, cross-modal structure progressively consolidates with depth around a compact set of recurrent hub neurons, whose targeted perturbation materially alters model output. Neural topology thus emerges as a meaningful intermediate scale for VLM interpretability: richer than local attribution, more tractable than full circuit recovery, and empirically tied to multimodal behavior. Code is publicly available at https://github.com/he-h/vlm-graph-probing.

Symmetry detection is a fundamental problem in computer vision, and symmetries serve as powerful priors for downstream tasks. However, existing learning-based methods for detecting 3D symmetries from single images have been almost exclusively trained and evaluated on object-centric or synthetic datasets, and thus fail to generalize to real-world scenes. Furthermore, due to the inherent scale ambiguity of monocular inputs, which makes localizing the 3D plane an ill-posed problem, many existing works only predict the plane's orientation. In this paper, we address these limitations by presenting the first framework for detecting *3D-grounded reflectional symmetries* from single, in-the-wild RGB images, focusing on architectural landmarks. We introduce two key innovations: (1) a scalable data annotation pipeline to automatically curate a large-scale dataset of architectural symmetries, ArchSym, from SfM reconstructions by leveraging cross-view image matching; and building on the dataset, (2) a single-view symmetry detector that accurately localizes symmetries in 3D by parameterizing them as signed distance maps defined relative to predicted scene geometry. We validate our symmetry annotation pipeline against geometry-based alternatives and demonstrate that our symmetry detector significantly outperforms state-of-the-art baselines on our new benchmark.

Despite the significant advancements in Large Vision-Language Models (LVLMs), their tendency to generate hallucinations undermines reliability and restricts broader practical deployment. Among the hallucination mitigation methods, feature steering emerges as a promising approach that reduces erroneous outputs in LVLMs without increasing inference costs. However, current methods apply uniform feature steering across all layers. This heuristic strategy ignores inter-layer differences, potentially disrupting layers unrelated to hallucinations and ultimately leading to performance degradation on general tasks. In this paper, we propose Locate-Then-Sparsify for Feature Steering (LTS-FS), a plug-and-play framework which controls the steering intensity according to the hallucination relevance of each layer. We first construct a dataset comprising token-level and sentence-level hallucination cases. Based on this dataset, we introduce an attribution method based on causal interventions to quantify the hallucination relevance of each layer. With the attribution scores across layers, we propose a layerwise strategy that converts these scores into feature steering intensities for individual layers, enabling more precise adjustments specifically on hallucination-relevant layers. Extensive experiments across multiple LVLMs and benchmarks demonstrate that LTS-FS effectively mitigates hallucination while preserving strong performance. Codes are available at https://github.com/huttersadan/LTS-FS.

As vision-language models are deployed at scale, understanding their internal mechanisms becomes increasingly critical. Existing interpretability methods predominantly rely on activations, making them dataset-dependent, vulnerable to data bias, and often restricted to coarse head-level explanations. We introduce SITH (Semantic Inspection of Transformer Heads), a fully data-free, training-free framework that directly analyzes CLIP's vision transformer in weight space. For each attention head, we decompose its value-output matrix into singular vectors and interpret each one via COMP (Coherent Orthogonal Matching Pursuit), a new algorithm that explains them as sparse, semantically coherent combinations of human-interpretable concepts. We show that SITH yields coherent, faithful intra-head explanations, validated through reconstruction fidelity and interpretability experiments. This allows us to use SITH for precise, interpretable weight-space model edits that amplify or suppress specific concepts, improving downstream performance without retraining. Furthermore, we use SITH to study model adaptation, showing how fine-tuning primarily reweights a stable semantic basis rather than learning entirely new features.

Video-based spatial reasoning -- such as estimating distances, judging directions, or understanding layouts from multiple views -- requires selecting informative frames and, when needed, actively seeking additional viewpoints during inference. Existing multimodal large language models (MLLMs) consume a fixed set of uniformly sampled frames and cannot request new views once reasoning begins, often missing the geometric cues necessary for reliable spatial judgments. We present EagleVision, a dual-stage framework that combines geometry-aware frame selection with active, Bird's-Eye-View (BEV)-grounded reasoning. In the first stage (macro perception), a semantics-perspective-fusion determinantal point process (SPF-DPP) selects a compact set of keyframes that jointly maximize semantic relevance and viewpoint diversity under a fixed token budget. In the second stage (micro verification), the model performs iterative spatial Chain-of-Thought: at each step it can either reason in text or predict a pose on the BEV plane to retrieve the nearest real frame, forming a closed-loop hypothesize-look-verify cycle. The querying policy is trained purely via reinforcement learning with a spatial grounding reward, requiring no human-annotated reasoning traces. On VSI-Bench and SQA3D, EagleVision achieves state-of-the-art performance among open-source vision-language models.