论文检索

Qinsong Li, Jing Meng, Haibo Wang, Shengjun Liu

Deep functional map frameworks (DFM) for shape correspondence are powerful, yet fundamentally limited by their reliance on end-to-end differentiability. This constraint prevents the integration of highly accurate, non-differentiable refinement techniques, capping their overall performance, especially on challenging non-isometric shapes. To overcome this, we introduce MDND, a novel DFM paradigm built on the principle of merging differentiable and non-differentiable components. Our framework facilitates unsupervised learning guided by an internal, non-differentiable refinement. Specifically, MDND employs a dual-branch architecture: a non-differentiable refinement branch leverages a novel, multiscale iterative solver to produce highly robust correspondences, acting as a refined target. Concurrently, a fully differentiable branch learns to predict correspondences from features. The entire system is trained end-to-end without supervision by enforcing a consistency loss that compels the differentiable branch to learn from the superior, refined results of the non-differentiable branch. Extensive experiments show that MDND sets a new state-of-the-art, demonstrating remarkable robustness on shapes with non-isometric deformations and topological noise.

Mingxuan Li, Zihao Huang, Xiaohui Chu, Fazhan Zhang, Bohan Fu, Runze Hu

Vision-Language Models (VLMs) have demonstrated significant progress in quality assessment tasks. However, a fundamental paradox arises when their application to Point Cloud Quality Assessment (PCQA). Existing VLMs, designed for image-text pairs, are inherently incompatible with 3D point cloud data due to the modality gap. While some PCQA research attempts to adapt point clouds to VLMs by 2D projection, this approach inevitably sacrifices crucial spatial structure information essential for accurate quality assessment. Conversely, directly integrating a dedicated 3D branch into a VLM-based PCQA framework introduces feature space misalignment and an influx of quality-insensitive information. To bridge these fundamental conflicts hindering VLMs' adaptation to PCQA, we propose the PMP-PCQA framework, which leverages the inherent mapping relationship between points and pixels to seamlessly apply VLMs to PCQA. Our approach introduces three key innovations: a Spatial Awareness Enhancer(SAE) module that enriches the image features with spatial coordinate clues to reinforce geometric awareness in 2D visual representations; a Fine-to-coarse Consistency Alignment(FCA) module that bridges the gap between 2D and 3D modalities by leveraging point-pixel correspondences to construct bridging features; and a Text-Guided Adaptive Miner(TAM) module that dynamically suppresses quality-insensitive features to mine discriminative visual clues for PCQA. Extensive evaluations demonstrate that PMP-PCQA consistently outperforms state-of-the-art methods across multiple benchmarks.

Mingxuan Li, Fazhan Zhang, Zhenzhe Hou, Zihao Huang, Bohan Fu, Runze Hu, Xiaohui Chu

Point Cloud Quality Assessment (PCQA) faces a critical disconnect: existing methods operate on a flawed single-perception paradigm, while human observers evaluate quality through dual cognitive streams: technical rationality and semantic sensibility. This fundamental mismatch routinely produces assessment failures in real-world scenarios where technical and semantic signals conflict. To address this, we introduce Dual-Stream Perception PCQA (DSP-PCQA), the first framework that explicitly models this perceptual duality through parallel networks thoroughly mirroring the human cognitive pathway. DSP-PCQA introduces three key innovations: (1) a Decoupled Focus Enhancer (DFE) that surgically isolates technical and semantic information using two targeted transformations; (2) a Context & Attribute Correlation Awareness (CACA) module that captures the dynamic, non-linear relationships between different views and sub-models characteristic of human visual processing; and (3) an Exchange-based Perceptual Injection (EPI) module that strategically transfers information between perception streams, simulating how humans integrate multiple perceptual dimensions. Extensive evaluations show DSP-PCQA outperforms state-of-the-art methods across multiple benchmarks. Most importantly, our method resolves the perceptual discord that plagues existing approaches, maintaining high accuracy even in the challenging boundary cases where technical quality and semantic significance diverge, precisely where conventional methods often struggle.

Liwen Li, Xinrui Guo, Wentao Guo, Shunqi Yang, Fumin Guo

Deformable medical image registration is essential in medical image analyses. Recent transformer-based registration methods have achieved high registration accuracy. However, these methods often rely on patch embedding at the beginning of encoding, resulting in limited ability to capture detailed anatomical structural information in the images and explore local semantic relationships within individual patches. Here, we proposed a novel Dual-feet Encoder (DFEnc) to asynchronously model semantic information from moving and fixed images at various scales through two separate branches in three steps. For each step, features from adjacent resolution levels were processed by a Single Step Hybrid Extractor (SSHExt), which performed patch convolution to preserve local information, followed by several transformer blocks to capture global context. Dense connections were employed to enhance semantic awareness across adjacent feature resolution levels. Additionally, we introduced a Feature Fusion-based Decoder (FFDec) to progressively fuse features related to the fixed and moving images and to generate intermediate deformation fields at each stage, enabling accurate image alignment through stepwise warping and alignment refinement. Extensive ablation studies demonstrated the effectiveness of the proposed DFEnc, SSHExt, and FFDec. Compared to a state-of-the-art AutoFuse-Trans method, our approach yielded improvements in Dice of 1.14%, 1.77%, and 4.47% on the ACDC, OASIS, and Abdomen CT datasets, respectively, while maintaining relatively low computational cost. These results suggest the utility of the proposed approach for broad research and clinical applications.

Linfei Li, Lin Zhang, Zhong Wang, Ying Shen

Recent advances in generative AI have accelerated the production of ultra-high-resolution visual content. However, traditional image formats face significant limitations in efficient compression and real-time decoding, which restricts their applicability on end-user devices. Inspired by 3D Gaussian Splatting, 2D Gaussian image models have achieved notable progress in enhancing image representation efficiency and quality. Nevertheless, existing methods struggle to balance compression ratios and reconstruction fidelity in ultra-high-resolution scenarios. To address these challenges, we propose SmartSplat, a highly adaptive and feature-aware GS-based image compression framework that effectively supports arbitrary image resolutions and compression ratios. By leveraging image-aware features such as gradients and color variances, SmartSplat introduces a Gradient-Color Guided Variational Sampling strategy alongside an Exclusion-based Uniform Sampling scheme, significantly improving the non-overlapping coverage of Gaussian primitives in pixel space. Additionally, a Scale-Adaptive Gaussian Color Sampling method is proposed to enhance the initialization of Gaussian color attributes across scales. Through joint optimization of spatial layout, scale, and color initialization, SmartSplat can efficiently capture both local structures and global textures of images using a limited number of Gaussians, achieving superior reconstruction quality under high compression ratios. Extensive experiments on DIV8K and a newly created 16K dataset demonstrate that SmartSplat significantly outperforms state-of-the-art methods at comparable compression ratios and surpasses their compression limits, exhibiting strong scalability and practical applicability. This framework can effectively alleviate the storage and transmission burdens of ultra-high-resolution images, providing a robust foundation for future high-efficiency visual content processing.

Lin Li, Wei Chen, Jiahui Li, Kwang-Ting Cheng, Long Chen

Recent advances in multi-modal large language models (MLLMs) have significantly improved object-level grounding and region captioning. However, they remain limited in visual relation understanding, struggling even with binary relation detection, let alone N-ary relations involving multiple semantic roles. The core reason is the lack of modeling for structural semantic dependencies among multi-entities, leading to over-reliance on language priors (e.g., defaulting to "person drinks a milk" if a person is merely holding it). To this end, we propose Relation-R1, the first unified relation comprehension framework that explicitly integrates cognitive chain-of-thought (CoT)-guided supervised fine-tuning (SFT) and group relative policy optimization (GRPO) within a reinforcement learning (RL) paradigm. Specifically, we first establish foundational reasoning capabilities via SFT, enforcing structured outputs with thinking processes. Then, GRPO is utilized to refine these outputs via multi-rewards optimization, prioritizing visual-semantic grounding over language-induced biases, thereby improving generalization capability. Furthermore, we investigate the impact of various CoT strategies within this framework, demonstrating that a specific-to-general progressive approach in CoT guidance further improves generalization, especially in capturing synonymous N-ary relations. Extensive experiments on widely-used PSG and SWiG datasets demonstrate that Relation-R1 achieves state-of-the-art performance in both binary and N-ary relation understanding.

Lei Li, Sen Jia, Jenq-Neng Hwang

We introduce LLaMMo (Large Language and Multi-Person Motion Assistant), the first instruction-tuning multimodal framework tailored for multi-human motion analysis. LLaMMo incorporates a novel human-centric and social-temporal learner that models and fuses both intra-person dynamics and inter-person dependencies, yielding robust, context-aware representations of complex group behaviors while maintaining low computational overhead. To support LLaMMo, we construct LLaVerse, a large-scale dataset with fine-grained manual annotations covering diverse multi-person activities spanning daily social interaction and professional team sports. Built on top of LLaVerse, we also propose LLaMI-Bench, a dedicated benchmark for evaluating multi-human behavior understanding across motion and video modalities. Extensive experiments demonstrate that LLaMMo consistently outperforms baselines in understanding multi-person interactions under low-latency settings, with notable gains in both social and sport-specific contexts.

Ke Li, Di Wang, Ting Wang, Fuyu Dong, Yiming Zhang, Luyao Zhang, Xiangyu Wang, Shaofeng Li, Quan Wang

Remote sensing visual grounding (RSVG) aims to localize objects in remote sensing images based on free-form natural language expressions. Existing approaches are typically constrained to closed-set vocabularies, limiting their applicability in open-world scenarios. While recent attempts to leverage generic foundation models for open-vocabulary RSVG, they overly rely on expensive high-quality datasets and time-consuming fine-tuning. To address these limitations, we propose RSVG-ZeroOV, a training-free framework that aims to explore the potential of frozen generic foundation models for zero-shot open-vocabulary RSVG. Specifically, RSVG-ZeroOV comprises three key stages: (i) Overview: We utilize a vision-language model (VLM) to obtain cross-attention maps that capture semantic correlations between text queries and visual regions. (ii) Focus: By leveraging the fine-grained modeling priors of a diffusion model (DM), we fill in gaps in structural and shape information of objects, which are often overlooked by VLM. (iii) Evolve: A simple yet effective attention evolution module is introduced to suppress irrelevant activations, yielding purified segmentation masks over the referred objects. Without cumbersome task-specific training, RSVG-ZeroOV offers an efficient and scalable solution. Extensive experiments demonstrate that the proposed framework consistently outperforms existing weakly-supervised and zero-shot methods.

Kaiyu Li, Xiangyong Cao, Yupeng Deng, Chao Pang, Zepeng Xin, Hui Qiao, Tieliang Gong, Deyu Meng, Zhi Wang

Monitoring Earth's evolving land covers requires methods capable of detecting changes across a wide range of categories and contexts. Existing change detection methods are hindered by their dependency on predefined classes, reducing their effectiveness in open-world applications. To address this issue, we introduce open-vocabulary change detection (OVCD), a novel task that bridges vision and language to detect changes across any category. Considering the lack of high-quality data and annotation, we propose two training-free frameworks, M-C-I and I-M-C, which leverage and integrate off-the-shelf foundation models for the OVCD task. The insight behind the M-C-I~framework is to discover all potential changes and then classify these changes, while the insight of I-M-C~framework is to identify all targets of interest and then determine whether their states have changed. Based on these two frameworks, we instantiate to obtain several methods, e.g., SAM-DINOv2-SegEarth-OV, Grounding-DINO-SAM2-DINO, etc. Extensive evaluations on 4 benchmark datasets demonstrate the superior generalization and robustness of our OVCD methods over existing supervised and unsupervised methods. To support continued exploration, we release DynamicEarth, a dedicated codebase designed to advance research and application of OVCD.

Kai Li, Wei Wang, Linchao Zhang, Siying Zhu, Wenqi Ren

The widespread and inconsistent compression applied by Online Social Networks severely degrades the performance of synthetic image detectors. We attribute this degradation to two main issues: 1) the model confuses forgery artifacts with compression artifacts, and 2) compression erodes crucial discriminative high-frequency details. Existing methods suppress compression features during training but overlook the overlap between compression features and forgery-related features, leading to the unintended removal of forgery traces. To address artifact confusion, we introduce a Decision-Driven Orthogonal Constraint, which defines a classification decision axis pointing from the real class centroid to the forged class centroid. This constraint enforces compression artifacts to be orthogonal to the decision axis, mitigating their interference with forgery detection without entirely removing them, thus preventing the suppression of forgery-related features. To mitigate the erosion of high-frequency details, we propose to mine complementary forgery cues from both low-frequency information and compressed high-frequency components. A bidirectional update strategy and an adaptive global-local modulator are proposed to facilitate the utilization of forgery cues. Extensive experiments demonstrate that our method achieves state-of-the-art generalization performance in challenging open-world detection scenarios.

Jun Li, Zikun Chen, Haibo Chen, Shuo Chen, Jian Yang

Novel object synthesis by integrating distinct textual concepts from diverse categories remains a significant challenge in text-to-image generation. Existing methods often suffer from insufficient concept mixing, lack of rigorous evaluation, and suboptimal outputs, resulting in conceptual imbalance, superficial combinations, or mere juxtapositions. To address these limitations, we propose Reinforcement Mixing Learning (RMLer), a framework that formulates cross-category concept fusion as a reinforcement learning problem: mixed features serve as states, mixing strategies as actions, and visual outcomes as rewards. Specifically, we design an MLP policy network to predict dynamic coefficients for blending cross-category text embeddings. We further introduce visual rewards based on (1) semantic similarity and (2) compositional balance between the fused object and its constituent concepts, and optimize the policy via proximal policy optimization. At inference time, a selection strategy leverages these rewards to curate the highest-quality fused objects. Extensive experiments demonstrate that RMLer synthesizes coherent, high-fidelity objects from diverse categories and consistently outperforms existing methods. Our work provides a robust framework for generating novel visual concepts, with promising applications in film, gaming, and design.

Jinxuan Li, Yi Zhang, Jian-Fang Hu, Chaolei Tan, Tianming Liang, Beihao Xia

Spatio-Temporal Video Grounding (STVG) aims to localize a spatio-temporal tube that corresponds to a given language query in an untrimmed video. This is a challenging task since it involves complex vision-language understanding and spatiotemporal reasoning. Recent works have explored weakly-supervised setting in STVG to eliminate reliance on fine-grained annotations like bounding boxes or temporal stamps. However, they typically follow a simple late-fusion manner, which generates tubes independent of the text description, often resulting in failed target identification and inconsistent target tracking. To address this limitation, we propose a Tube-conditioned Reconstruction with Mutual Constraints (TubeRMC) framework that generates text-conditioned candidate tubes with pre-trained visual grounding models and further refine them via tube-conditioned reconstruction with spatio-temporal constraints. Specifically, we design three reconstruction strategies from temporal, spatial, and spatio-temporal perspectives to comprehensively capture rich tube-text correspondences. Each strategy is equipped with a Tube-conditioned Reconstructor, utilizing spatio-temporal tubes as condition to reconstruct the key clues in the query. We further introduce mutual constraints between spatial and temporal proposals to enhance their quality for reconstruction. TubeRMC outperforms existing methods on two public benchmarks VidSTG and HCSTVG. Further visualization shows that TubeRMC effectively mitigates both target identification errors and inconsistent tracking.

Jingyao Li, Jingyun Wang, Molin Tan, Haochen Wang, Cilin Yan, Likun Shi, Jiayin Cai, Xiaolong Jiang, Yao Hu

Cross-Video Reasoning (CVR) presents a significant challenge in video understanding, which requires simultaneous understanding of multiple videos to aggregate and compare information across groups of videos. Most existing video understanding benchmarks focus on single-video analysis, failing to assess the ability of multimodal large language models (MLLMs) to simultaneously reason over various videos. Recent benchmarks evaluate MLLMs' capabilities on multi-view videos that capture different perspectives of the same scene. However, their limited tasks hinder a thorough assessment of MLLMs in diverse real-world CVR scenarios. To this end, we introduce CrossVid, the first benchmark designed to comprehensively evaluate MLLMs' spatial-temporal reasoning ability in cross-video contexts. Firstly, CrossVid encompasses a wide spectrum of hierarchical tasks, comprising four high-level dimensions and ten specific tasks, thereby closely reflecting the complex and varied nature of real-world video understanding. Secondly, CrossVid provides 5,331 videos, along with 9,015 challenging question-answering pairs, spanning single-choice, multiple-choice, and open-ended question formats. Through extensive experiments on various open-source and closed-source MLLMs, we observe that Gemini-2.5-Pro performs best on CrossVid, achieving an average accuracy of 50.4%. Notably, our in-depth case study demonstrates that most current MLLMs struggle with CVR tasks, primarily due to their inability to integrate or compare evidence distributed across multiple videos for reasoning. These insights highlight the potential of CrossVid to guide future advancements in enhancing MLLMs’ CVR capabilities.

Jingxiong Li, Chenglu Zhu, Sunyi Zheng, Yuxuan Sun, Yifei Wang, He Liu, Yunlong Zhang, Yixuan Si, Lin Yang, Liang Xiao

We present MIRA (Multimodal Interventional RAdiology evaluation), a comprehensive benchmark for evaluating large multimodal models in expert-level interventional radiology tasks requiring specialized domain knowledge and advanced visual reasoning capabilities. Unlike existing medical benchmarks that primarily provide binary labels without contextual depth, MIRA offers diverse question formats, including open-ended, closed-ended, single-choice, and multiple-choice categories, each accompanied by detailed expert-validated explanations. The benchmark incorporates approximately 184K high-quality medical images spanning multiple imaging modalities with 1.2M meticulously generated question-answer pairs across various anatomical regions. These pairs were created through a sophisticated cascade methodology involving expert interventional radiologists at both the data collection and validation stages. Our comprehensive evaluation, encompassing zero-shot testing and fine-tuning experiments of large multimodal models, revealing significant performance gaps between AI systems and human specialists. Fine-tuning experiments demonstrate substantial improvements, with models achieving up to 0.80 accuracy on single-choice questions. MIRA establishes a challenging benchmark that suggests promising directions for developing specialized clinical AI systems for interventional radiology.

Jing Li, Yifan Wang, Jiafeng Yan, Renlong Zhang, Bin Yang

Infrared and visible image fusion aims to integrate complementary multi-modal information into a single fused result. However, existing methods 1) fail to account for the degradation visible images under adverse weather conditions, thereby compromising fusion performance; and 2) rely on fixed network architectures, limiting their adaptability to diverse degradation scenarios. To address these issues, we propose a one-stop degradation-aware image fusion framework for multi-degradation scenarios driven by a large language model (MdaIF). Given the distinct scattering characteristics of different degradation scenarios (e.g., haze, rain, and snow) in atmospheric transmission, a mixture-of-experts (MoE) system is introduced to tackle image fusion across multiple degradation scenarios. To adaptively extract diverse weather-aware degradation knowledge and scene feature representations, collectively referred to as the semantic prior, we employ a pre-trained vision-language model (VLM) in our framework. Guided by the semantic prior, we propose degradation-aware channel attention module (DCAM), which employ degradation prototype decomposition to facilitate multi-modal feature interaction in channel domain. In addition, to achieve effective expert routing, the semantic prior and channel-domain modulated features are utilized to guide the MoE, enabling robust image fusion in complex degradation scenarios. Extensive experiments validate the effectiveness of our MdaIF, demonstrating superior performance over SOTA methods.

Jie Li, Yingying Feng, Chi Xie, Jie Hu, Lei Tan, Jiayi Ji

The remarkable realism of images generated by diffusion models poses critical detection challenges. Current methods utilize reconstruction error as a discriminative feature, exploiting the observation that real images exhibit higher reconstruction errors when processed through diffusion models. However, these approaches require costly reconstruction computations and depend on specific diffusion models, making their performance highly model-dependent. We identify a fundamental difference: real images are more difficult to fit with Gaussian distributions compared to synthetic ones. In this paper, we propose Forgery Identification via Noise Disturbance (FIND), a novel method that requires only a simple binary classifier. It eliminates reconstruction by directly targeting the core distributional difference between real and synthetic images. Our key operation is to add Gaussian noise to real images during training and label these noisy versions as synthetic. This step allows the classifier to focus on the statistical patterns that distinguish real from synthetic images. We theoretically prove that the noise-augmented real images resemble diffusion-generated images in their ease of Gaussian fitting. Furthermore, simply by adding noise, they still retain visual similarity to the original images, highlighting the most discriminative distribution-related features. The proposed FIND improves performance by 11.7% on the GenImage benchmark while running 126x faster than existing methods. By removing the need for auxiliary diffusion models and reconstruction, it offers a practical, efficient, and generalizable way to detect diffusion-generated content.

Jiawen Li, Fei Jiang, Dandan Zhu, Jinxin Shi, Aimin Zhou

Unsupervised domain adaptive pose estimation is a fundamental yet challenging task due to the need to transfer from labeled synthetic data to unlabeled real data. Nevertheless, the underlying pose semantics, which are governed by spatial structure, remain largely consistent across domains. This observation motivates the use of vision-language models, which provide domain-invariant representations that align well with high-level semantic concepts. Motivated by this, we propose CLIP2Pose, a novel framework that leverages the semantic robustness of frozen CLIP encoders to facilitate cross-domain generalization. We first introduce a semantic-driven prompt mechanism that encodes structural priors, domain-specific appearance, and instance-level context into the image representation. This guides the model to focus on semantically meaningful and structurally relevant features. Next, we propose a semantic modulation module that adaptively refines visual features by conditioning them on prompt-derived embeddings, enhancing alignment between semantics and visual patterns. To further bridge the modality and domain gaps, we design a directional alignment loss that encourages consistent structural reasoning across both vision and language representations. Extensive experiments on domain adaptive human body and hand pose benchmarks show that CLIP2Pose achieves state-of-the-art performance.

Jiaoyang Li, Jun Fang, Tianhao Gao, Xiaohui Zhang, Zhiyuan Liu, Chao Liu, Pengzhang Liu, Qixia Jiang

Representation learning is fundamental to modern machine learning, powering applications such as text retrieval and multimodal understanding. However, learning robust and generalizable representations remains challenging. While prior work has demonstrated that active noise injection, a form of data augmentation, can enhance encoding performance, most existing methods rely on heuristic or static noise, overlooking the dynamic nature of feature distributions during training. In this work, we systematically study the role of noise in representation learning from both gradient-based and feature distribution perspectives, using InfoNCE loss as a representative example. Focusing on multimodal representation learning, we propose FANoise, a novel feature-adaptive noise injection strategy. By leveraging the dynamics of contrastive learning, FANoise effectively mitigates the negative impacts of noise while preserving its benefits. Under this theoretically grounded framework, comprehensive experiments demonstrate that FANoise consistently improves overall performance on multimodal tasks across various base VLM models.

Jiahao Li, Yusheng Luo, Yunzhong Lou, Xiangdong Zhou

We present ReCAD, a reinforcement learning (RL) framework that bootstraps pretrained large models (PLMs) to generate precise parametric computer-aided design (CAD) models from multimodal inputs by leveraging their inherent generative capabilities. With just access to simple functional interfaces (e.g., point coordinates), our approach enables the emergence of complex CAD operations (e.g., pattern replication and mirror). This stands in contrast to previous methods, which typically rely on knowledge injected through supervised fine-tuning (SFT), offer limited support for editability, and fail to exploit the strong generative priors of PLMs. Specifically, the ReCAD framework begins by fine-tuning vision-language models (VLMs) to equip them with basic CAD model generation capabilities, where we rewrite CAD scripts into parameterized code that is leveraged to generate accurate textual descriptions for supervision. Then, we propose a novel RL strategy that incorporates parameterized code as guidance to enhance the model’s reasoning on challenging questions. Furthermore, we employ a hierarchical primitive learning process to progressively teach structured and compositional skills under a unified reward function that ensures both geometric accuracy and semantic fidelity. ReCAD sets a new state-of-the-art in both text-to-CAD and image-to-CAD tasks, significantly improving geometric accuracy across in-distribution and out-of-distribution settings. In the image-to-CAD task, for instance, it reduces the mean Chamfer Distance from 73.47 to 29.61 (in-distribution) and from 272.06 to 80.23 (out-of-distribution), outperforming existing baselines by a substantial margin.

Jiahao Li, Yang Lu, Yachao Zhang, Yong Xie, Fangyong Wang, Yuan Xie, Yanyun Qu

Open-vocabulary semantic segmentation (OVSS) employs pixel-level vision-language alignment to associate category-related prompts with corresponding pixels. A key challenge is enhancing the multimodal dense prediction capability, specifically this pixel-level multimodal alignment. Although existing methods achieve promising results by leveraging CLIP’s vision-language alignment, they rarely investigate the performance boundaries of CLIP for dense prediction from an interpretability mechanisms perspective. In this work, we systematically investigate CLIP's internal mechanisms and identify a critical phenomenon: analogous to human distraction, CLIP diverts significant attention resources from target regions to irrelevant tokens. Our analysis reveals that these tokens arise from dimension-specific over-activation; filtering them enhances CLIP's dense prediction performance. Consequently, we propose Refocusing CLIP (RF-CLIP), a training-free approach that emulates human distraction-refocusing behavior to redirect attention from distraction tokens back to target regions, thereby refining CLIP's multimodal alignment granularity. Our method achieves SOTA performance on eight benchmarks while maintaining high inference efficiency.