论文检索

Tongshun Zhang, Pingping Liu, Zixuan Zhong, Zijian Zhang, Qiuzhan Zhou

Recovering fine-grained details in extremely dark images remains challenging due to severe structural information loss and noise corruption. Existing enhancement methods often fail to preserve intricate details and sharp edges, limiting their effectiveness in downstream applications like text and edge detection. To address these deficiencies, we propose an efficient dual-stage approach centered on detail recovery for dark images. In the first stage, we introduce a Residual Fourier-Guided Module (RFGM) that effectively restores global illumination in the frequency domain. RFGM captures inter-stage and inter-channel dependencies through residual connections, providing robust priors for high-fidelity frequency processing while mitigating error accumulation risks from unreliable priors. The second stage employs complementary Mamba modules specifically designed for textural structure refinement: (1) Patch Mamba operates on channel-concatenated non-downsampled patches, meticulously modeling pixel-level correlations to enhance fine-grained details without resolution loss. (2) Grad Mamba explicitly focuses on high-gradient regions, alleviating state decay in state space models and prioritizing reconstruction of sharp edges and boundaries. Extensive experiments on multiple benchmark datasets and downstream applications demonstrate that our method significantly improves detail recovery performance while maintaining efficiency. Crucially, the proposed modules are lightweight and can be seamlessly integrated into existing Fourier-based frameworks with minimal computational overhead.

Tianxiang Zhang, Peipeng Yu, Zhihua Xia, Longchen Dai, Xiaoyu Zhou, Hui Gao

The proliferation of sophisticated deepfakes poses significant threats to information integrity. While DINOv2 shows promise for detection, existing fine-tuning approaches treat it as generic binary classification, overlooking distinct artifacts inherent to different deepfake methods. To address this, we propose a DeepFake Fine-Grained Adapter (DFF-Adapter) for DINOv2. Our method incorporates lightweight multi-head LoRA modules into every transformer block, enabling efficient backbone adaptation. DFF-Adapter simultaneously addresses authenticity detection and fine-grained manipulation type classification, where classifying forgery methods enhances artifact sensitivity. We introduce a shared branch propagating fine-grained manipulation cues to the authenticity head. This enables multi-task cooperative optimization, explicitly enhancing authenticity discrimination with manipulation-specific knowledge. Utilizing only 3.5M trainable parameters, our parameter-efficient approach achieves detection accuracy comparable to or even surpassing that of current complex state-of-the-art methods.

Tianshuo Zhang, Siran Peng, Li Gao, Haoyuan Zhang, Xiangyu Zhu, Zhen Lei

The rapid advancements in face forgery techniques necessitate that detectors continuously adapt to new forgery methods, thus situating face forgery detection within a continual learning paradigm. However, when detectors learn new forgery types, their performance on previous types often degrades rapidly, a phenomenon known as catastrophic forgetting. Kolmogorov-Arnold Networks (KANs) utilize locally plastic splines as their activation functions, enabling them to learn new tasks by modifying only local regions of the functions while leaving other areas unaffected. Therefore, they are naturally suitable for addressing catastrophic forgetting. However, KANs have two significant limitations: 1) the splines are ineffective for modeling high-dimensional images, while alternative activation functions that are suitable for images lack the essential property of locality; 2) in continual learning, when features from different domains overlap, the mapping of different domains to distinct curve regions always collapses due to repeated modifications of the same regions. In this paper, we propose a KAN-based Continual Face Forgery Detection (KAN-CFD) framework, which includes a Domain-Group KAN Detector (DG-KD) and a data-free replay Feature Separation strategy via KAN Drift Compensation Projection (FS-KDCP). DG-KD enables KANs to fit high-dimensional image inputs while preserving locality and local plasticity. FS-KDCP avoids the overlap of the KAN input spaces without using data from prior tasks. Experimental results demonstrate that the proposed method achieves superior performance while notably reducing forgetting.

Tianlu Zhang, Qiang Zhang, Guiguang Ding, Jungong Han

Tracking and segmentation play essential roles in video understanding, providing basic positional information and temporal association of objects within video sequences. Despite their shared objective, existing approaches often tackle these tasks using specialized architectures or modality-specific parameters, limiting their generalization and scalability. Recent efforts have attempted to unify multiple tracking and segmentation sub-tasks from the perspectives of any modality input or multi-task inference. However, these approaches tend to overlook two critical challenges: the distributional gap across different modalities and the feature representation gap across tasks. These issues hinder effective cross-task and cross-modal knowledge sharing, ultimately constraining the development of a true generalist model. To address these limitations, we propose a universal tracking and segmentation framework named SATA, which unifies a broad spectrum of tracking and segmentation subtasks with any modality input. Specifically, a Decoupled Mixture-of-Expert (DeMoE) mechanism is presented to decouple the unified representation learning task into the modeling process of cross-modal shared knowledge and specific information, thus enabling the model to maintain flexibility while enhancing generalization. Additionally, we introduce a Task-aware Multi-object Tracking (TaMOT) pipeline to unify all the task outputs as a unified set of instances with calibrated ID information, thereby alleviating the degradation of task-specific knowledge during multi-task training. SATA demonstrates superior performance on 18 challenging tracking and segmentation benchmarks, offering a novel perspective for more generalizable video understanding.

Shuo Zhang, Jiaming Huang, Wenbing Tang, Jing Liu, LI HAN, Jiandun Li, Hongchun Yuan, Zizhu Fan

Multi-modal salient object detection (MSOD), which integrates complementary modalities such as depth or thermal data, primarily faces two challenges: accurately preserving salient object details and effectively aligning cross-modal features. Recent advances in using Stable Diffusion to generate images with fine edge details have inspired researchers to reformulate MSOD as a conditional mask generation process guided by salient features, which has achieved excellent visual results. However, these approaches often overlook the high computational cost and large-scale architecture of Stable Diffusion, both of which render it unsuitable for real-world MSOD applications. Therefore, we propose SimpleDiffusion, the first lightweight and efficient conditional diffusion model for MSOD that does not rely on Stable Diffusion. Specifically, we propose an Adaptive Cross-Modal Fusion Conditional Network and a Latent Denoising Network to reduce the complexity of diffusion models. Furthermore, we design a Multi-modal Feature Rectification and Fusion Module to enhance the representational capacity of cross-modal salient features. Customized training and sampling strategies are also developed to improve inference efficiency and reduce erroneous object segmentations. Experiments on multiple MSOD datasets demonstrate that SimpleDiffusion reduces model size by over tenfold and improves inference speed by more than fivefold compared to other diffusion-based methods, while maintaining comparable or superior performance.

Shizhou Zhang, Xueqiang Lv, Yinghui Xing, Qirui Wu, Di Xu, Chen Zhao, Yanning Zhang

Current methodologies for incremental object detection (IOD) primarily rely on Faster R-CNN or DETR series detectors; however, these approaches do not accommodate the real-time YOLO detection frameworks. In this paper, we first identify three primary types of knowledge conflicts that contribute to catastrophic forgetting in YOLO-based incremental detectors: foreground-background confusion, parameter interference, and misaligned knowledge distillation. Subsequently, we introduce YOLO-IOD, a real-time Incremental Object Detection (IOD) framework that is constructed upon the pretrained YOLO-World model, facilitating incremental learning via a stage-wise parameter-efficient finetuning process. Specifically, YOLO-IOD encompasses three principal components: 1) Conflict-Aware Pseudo-Label Refinement (CPR), which mitigates the foreground-background confusion by leveraging the confidence levels of pseudo labels and identifying potential objects relevant to future tasks. 2) Importance-based Kernel Selection (IKS), which identifies and updates the pivotal convolution kernels pertinent to the current task during the current learning stage. 3)Cross-Stage Asymmetric Knowledge Distillation (CAKD), which addresses the misaligned knowledge distillation conflict by transmitting the features of the student target detector through the detection heads of both the previous and current teacher detectors, thereby facilitating asymmetric distillation between existing and newly introduced categories. We further introduce LoCo COCO, a more realistic benchmark that eliminates data leakage across stages. Experiments on both conventional and LoCo COCO benchmarks show that YOLO-IOD achieves superior performance with minimal forgetting.

Shilong Zhang, Wenbo Li, Shoufa Chen, Chongjian GE, Peize Sun, Yifu Zhang, Yi Jiang, Zehuan Yuan, Bingyue Peng, Ping Luo

DiT models have achieved great success in text-to-video generation, leveraging their scalability in model capacity and data scale. High content and motion fidelity aligned with text prompts, however, often require large model parameters and a substantial number of function evaluations (NFEs). Realistic and visually appealing details are typically reflected in high-resolution outputs, further amplifying computational demands—especially for single-stage DiT models. To address these challenges, we propose a novel two-stage framework, FlashVideo, which strategically allocates model capacity and NFEs across stages to balance generation fidelity and quality. In the first stage, prompt fidelity is prioritized through a low-resolution generation process utilizing large parameters and sufficient NFEs to enhance computational efficiency. The second stage achieves a nearly straight ODE trajectory between low and high resolutions via flow matching, effectively generating fine details and fixing artifacts with minimal NFEs. To ensure a seamless connection between the two independently trained stages during inference, we carefully design degradation strategies during the second-stage training. Quantitative and visual results demonstrate that FlashVideo achieves state-of-the-art high-resolution video generation with superior computational efficiency. Additionally, the two-stage design enables users to preview the initial output and accordingly adjust the prompt before committing to full-resolution generation, thereby significantly reducing computational costs and wait times as well as enhancing commercial viability.

Shi-Xue Zhang, Hongfa Wang, Duojun Huang, Xin Li, Xiaobin Zhu, Xu-Cheng Yin

Video captions play a crucial role in text-to-video generation tasks, as their quality directly influences the semantic coherence and visual fidelity of the generated videos. Although large vision-language models (VLMs) have demonstrated significant potential in caption generation, existing benchmarks inadequately address fine-grained evaluation, particularly in capturing spatial-temporal details critical for video generation. To address this gap, we introduce the Fine-grained Video Caption Evaluation Benchmark (VCapsBench), the first large-scale fine-grained benchmark comprising 5,677 (5K+) videos and 109,796 (100K+) question-answer pairs. These QA-pairs are systematically annotated across 21 fine-grained dimensions (e.g., camera movement, and shot type) that are empirically proven critical for text-to-video generation. We further introduce three metrics (Accuracy (AR), Inconsistency Rate (IR), Coverage Rate (CR)), and an automated evaluation pipeline leveraging a large language model (LLM) to verify caption quality via contrastive QA-pairs analysis. Our benchmark can advance the development of robust text-to-video models by providing actionable insights for caption optimization.

Ruiqiang Zhang, Zehua Ma, Guanjie Wang, Chang Liu, Hengyi Wang, Weiming Zhang

With the deepening trend of paperless workflows, signatures as a means of identity authentication are gradually shifting from traditional ink-on-paper to electronic formats. Despite the availability of dynamic pressure-sensitive and PKI-based digital signatures, static scanned signatures remain prevalent in practice due to their convenience. However, these static images, having almost lost their authentication attributes, cannot be reliably verified and are vulnerable to malicious copying and reuse. To address these issues, we propose AuthSig, a novel static electronic signature framework based on generative models and watermark, which binds authentication information to the signature image. Leveraging the human visual system’s insensitivity to subtle style variations, AuthSig finely modulates style embeddings during generation to implicitly encode watermark bits-enforcing a One Signature, One Use policy. To overcome the scarcity of handwritten signature data and the limitations of traditional augmentation methods, we introduce a keypoint-driven data augmentation strategy that effectively enhances style diversity to support robust watermark embedding. Experimental results show that AuthSig achieves over 98% extraction accuracy under both digital-domain distortions and signature-specific degradations, and remains effective even in print-scan scenarios.

Ruicheng Zhang, Jun Zhou, Zunnan Xu, Zihao Liu, Jiehui Huang, Mingyang Zhang, Yu Sun, Xiu Li

Trajectory-Guided image-to-video (I2V) generation aims to synthesize videos that adhere to user-specified motion instructions. Existing methods typically rely on computationally expensive fine-tuning on scarce annotated datasets. Although some zero-shot methods attempt to trajectory control in the latent space, they may yield unrealistic motion by neglecting 3D perspective and creating a misalignment between the manipulated latents and the network's noise predictions. To address these challenges, we introduce Zo3T, a novel zero-shot test-time-training framework for trajectory-guided generation with three core innovations: First, we incorporate a 3D-Aware Kinematic Projection, leveraging inferring scene depth to derive perspective-correct affine transformations for target regions. Second, we introduce Trajectory-Guided Test-Time LoRA, a mechanism that dynamically injects and optimizes ephemeral LoRA adapters into the denoising network alongside the latent state. Driven by a regional feature consistency loss, this co-adaptation effectively enforces motion constraints while allowing the pre-trained model to locally adapt its internal representations to the manipulated latent, thereby ensuring generative fidelity and on-manifold adherence. Finally, we develop Guidance Field Rectification, which refines the denoising evolutionary path by optimizing the conditional guidance field through a one-step lookahead strategy, ensuring efficient generative progression towards the target trajectory. Zo3T significantly enhances 3D realism and motion accuracy in trajectory-controlled I2V generation, demonstrating superior performance over existing training-based and zero-shot approaches.

Rong Zhang, Jinxiao Li, Jingnan Wang, Zhiwen Zuo, Jianfeng Dong, Wei Li, Chi Wang, Weiwei Xu, Xun Wang

Garment-centric fashion image generation aims to synthesize realistic and controllable human models dressing a given garment, which has attracted growing interest due to its practical applications in e-commerce. The key challenges of the task lie in two aspects: (1) faithfully preserving the garment details, and (2) gaining fine-grained controllability over the model's appearance. Existing methods typically require performing garment deformation in the generation process, which often leads to garment texture distortions. Also, they fail to control the fine-grained attributes of the generated models, due to the lack of specifically designed mechanisms. To address these issues, we propose FashionMAC, a novel diffusion-based deformation-free framework that achieves high-quality and controllable fashion showcase image generation. The core idea of our framework is to eliminate the need for performing garment deformation and directly outpaint the garment segmented from a dressed person, which enables faithful preservation of the intricate garment details. Moreover, we propose a novel region-adaptive decoupled attention (RADA) mechanism along with a chained mask injection strategy to achieve fine-grained appearance controllability over the synthesized human models. Specifically, RADA adaptively predicts the generated regions for each fine-grained text attribute and enforces the text attribute to focus on the predicted regions by a chained mask injection strategy, significantly enhancing the visual fidelity and the controllability. Extensive experiments validate the superior performance of our framework compared to existing state-of-the-art methods.

Qin Zhang, Kun Zhou, Xulun Ye

3D full-scene segmentation technology has demonstrated great potential driven by large models, but it often faces challenges of incomplete scenes and identification of invisible classes in practical applications. To address this, we propose the LR-AdaInSeg method, which significantly enhances the model’s generalization ability in incomplete scenes through two key innovations: First, we design a Bayesian Low-Rank Module, which effectively solves the problem of feature space redundancy through dynamic optimization of the network structure, improving adaptability to incomplete scenes. Second, we combine graph contrastive clustering with the Low-Rank module, leveraging its robust feature representation capability to achieve accurate differentiation of invisible classes. In terms of implementation, we build a multi-scale feature extraction framework based on the 3D U-Net and utilize the 3D prompt points and their 2D masks as supervisory signals to achieve effective fusion of geometric and semantic information. Experiments show that our method achieves advanced performance on multiple benchmarks such as ScanNet, particularly excelling in handling incomplete scenes and invisible class objects.

Peng Zhang, Zhihui Lai, Wenting Chen, Xu Wu, Heng Kong

Medical vision-language pre-training (VLP) offers significant potential for advancing medical image understanding by leveraging paired image-report data. However, existing methods are limited by False Negatives (FaNe) induced by semantically similar texts and insufficient fine-grained cross-modal alignment. To address these limitations, we propose FaNe, a semantic-enhanced VLP framework. To mitigate false negatives, we introduce a semantic-aware positive pair mining strategy based on text-text similarity with adaptive normalization. Furthermore, we design a text-conditioned sparse attention pooling module to enable fine-grained image-text alignment through localized visual representations guided by textual cues. To strengthen intra-modal discrimination, we develop a hard-negative aware contrastive loss that adaptively reweights semantically similar negatives. Extensive experiments on five downstream medical imaging benchmarks demonstrate that FaNe achieves state-of-the-art performance across image classification, object detection, and semantic segmentation, validating the effectiveness of our framework.

Peng Zhang, Xiang Yuan, Gong Cheng

Domain generalization remains a critical challenge for deploying neural networks, particularly in out-of-distribution object detection. The distributional discrepancy between training (e.g., daytime-sunny) and the realistic condition (e.g., night-rainy) inevitably produces imprecise localization and wrong classification. To address these issues, we propose a unified interaction consistency learning (UICL) framework, a novel single-source domain-generalized method designed to learn intra-class domain-invariant representations. Specifically, we put forth a cross-domain interaction mechanism to exchange region proposals between original and augmented pipelines, enriching the diversity of instance-level representations. Building upon this, we propose prediction-guided consistency learning to unify the interaction mechanism and harmonize the cross-domain representations, contributing to a discriminative prediction distribution under domain shift. In addition, we devise a cyclic interaction resilient detection strategy, which mitigates inaccurate predictions suffering from partial occlusion and ambiguous boundaries among different domains. Extensive experiments evidence that UICL significantly improves the robustness of detectors over several target domains, achieving state-of-the-art generalization performance on the diverse weather benchmark.

Mingwei Zhang, Jingtao Hu, Qiang Li, Qi Wang

Remote sensing change detection (CD) has achieved remarkable progress in recent years. However, little attention has been paid to generalizable change detection (GCD) methods that can effectively generalize to unseen scenarios or domains beyond the training distribution. The major challenges in GCD arise from domain diversity and bitemporal domain shifts in remote sensing images, caused by variations in imaging platforms, acquisition times, geographic regions, and observed events. To tackle these challenges, we propose GenCD, a GCD framework built upon vision foundation models (VFMs). Specifically, GenCD introduces two key components: (1) a Low-Rank Exchange Adaptation (LREA) strategy of VFMs that aligns bitemporal representations while preserving the generalization capacity of VFMs on single-temporal inputs; and (2) a Token-Guided Feature Refinement (TGFR) mechanism that leverages an input-independent token as a guide to refine difference features, improving the discrimination between changed and unchanged regions. We conduct extensive cross-dataset evaluations on eight diverse datasets across three binary CD tasks: land cover, land use, and building-only CD. The results consistently demonstrate the superior generalization of GenCD over SoTA methods, highlighting its effectiveness in GCD.

Mingjin Zhang, Zhongkai Yang, Fei Gao

Due to the difficulties of directly obtaining high-resolution hyperspectral images (HR-HSI), the fusion of low-resolution hyperspectral images (LR-HSI) and high-resolution multispectral images (HR-MSI) has emerged as an effective approach. While existing methods leverage image-level priors from HR-MSI, they often lack explicit semantic guidance for precise detail reconstruction. Recognizing that textual scene descriptions encapsulate valuable object attributes and contextual information, we introduce the first Language-Bridging framework for Hyperspectral and Multispectral image fusion (CO²IF). CO²IF leverages language semantics as prior knowledge to explicitly guide the reconstruction process. To bridge the modality gap between textual descriptions and high-dimensional hyperspectral data, we design a Cross-modal Optimal Transport (COT) module. COT establishes precise semantic correspondences between language features and the visual cues of individual spectral bands. Building upon this semantic alignment, we develop a Multimodal Coordinated State Space Model (CoMamba). CoMamba effectively integrates the language-derived priors with spatial information from HR-MSI and spectral information from LR-HSI. This language-guided reconstruction significantly enhances the extraction of crucial spatial-spectral details, leading to superior fidelity in the generated HR-HSI. In addition, this paper adds text descriptions for three widely used datasets. Both qualitative and quantitative experimental results on the public datasets confirm the superiority of the proposed method compared to the SOTA methods.

Liwen Zhang, Zhaoji Miao, Yusong Shen, Zechen Wei, Hui Hui, Jie Tian

Magnetic Particle Imaging (MPI) is an innovative medical modality, providing nanomolar-scale in vivo sensitivity and radiation-free dynamic real-time detection for precision medicine. However, MPI faces a challenging problem in accurately visualizing nanoparticle distributions, where the reconstructed images with unidirectional scanning exhibit anisotropy. The anisotropy in spatial resolution leads to distortion and blurred image boundaries. Existing deep learning methods for anisotropy calibration are only limited to simulation data due to lacking of real-world MPI datasets. To address the aforementioned problems, we spent over three years designing and constructing a real-world MPI anisotropic image datasets (20,156 images) with diverse phantoms (sensitivity, resolution, vessel, shape) and animal scanning. Then, we introduce a novel Mamba-based method, MPI-Mamba, for anisotropic image calibration. Specifically, we propose a latent feature fusion state space model (LFF-SSM) block for feature fusion and leverage conditional latent diffusion model (CL-DM) branch for feature extraction. The CL-DM is performed to extract latent features in a highly compressed latent space for guiding the calibration and deblurring process. Next, we exploit the LFF-SSM to fully fuse the extracted multi-scale features to capture contextual information from the image structure, enabling the model to learn the overall distribution of signal concentration. We evaluate our method and competing methods on simulation dataset and our constructed diverse real-world MPI datasets. The results show that our proposed approach outperforms competing methods for anisotropic image calibration and deblurring.

Lirui Zhang, Zhengkai Zhao, Zhi Zuo, Pan Gao, Jie Qin

Point cloud completion is a fundamental task in 3D vision. A persistent challenge in this field is simultaneously preserving fine-grained details present in the input while ensuring the global structural integrity of the completed shape. While recent works leveraging local symmetry transformations via direct regression have significantly improved the preservation of geometric structure details, these methods suffer from two major limitations: (1) These regression-based methods are prone to overfitting which tend to memorize instant-specific transformations instead of learning a generalizable geometric prior. (2) Their reliance on point-wise transformation regression lead to high sensitivity to input noise, severely degrading their robustness and generalization. To address these challenges, we introduce Simba, a novel framework that reformulates point-wise transformation regression as a distribution learning problem. Our approach integrates symmetry priors with the powerful generative capabilities of diffusion models, avoiding instance-specific memorization while capturing robust geometric structures. Additionally, we introduce a hierarchical Mamba-based architecture to achieve high-fidelity upsampling. Extensive experiments across the PCN, ShapeNet, and KITTI benchmarks validate our method's state-of-the-art (SOTA) performance.

Lingfeng Zhang, Haoxiang Fu, Xiaoshuai Hao, Shuyi Zhang, Qiang Zhang, Rui Liu, Long Chen, Wenbo Ding

Embodied navigation is a fundamental capability that enables embodied agents to effectively interact with the physical world in various complex environments. However, a significant gap remains between current embodied navigation tasks and real-world requirements, as existing methods often struggle to integrate high-level human instructions with spatial understanding. To address this gap, we propose a new task of embodied navigation called spatial navigation, which encompasses two key components: spatial object navigation (SpON) for object-specific guidance and spatial area navigation (SpAN) for navigating to designated areas. Specifically, SpON guides agents to specific objects by leveraging spatial relationships and contextual understanding, while SpAN focuses on navigating to defined areas within complex environments. Together, these components significantly enhance agents’ navigation capabilities, enabling more effective interactions in real-world scenarios. To support this task, we have generated a spatial navigation dataset consisting of 10K trajectories within the simulator. This dataset includes high-level human instructions, detailed observations, and corresponding navigation actions, providing a comprehensive resource to enhance agent training and performance. Building on the spatial navigation dataset, we introduce SpNav, a hierarchical navigation framework. Specifically, SpNav employs vision-language model (VLM) to interpret high-level human instructions and accurately identify goal objects or areas within the observation range, achieving precise point-to-point navigation using a map and enhancing the agent’s ability to oper- ate effectively in complex environments by bridging the gap between perception and action. Extensive experiments show that SpNav achieves state-of-the-art (SOTA) performance in spatial navigation tasks across both simulated and real-world environments, validating the effectiveness of our method.

Liheng Zhang, Jin Wang, Hui Li, Bingfeng Zhang, Weifeng Liu

3D understanding has drawn significant attention recently, leveraging Vision-Language Models (VLMs) to enable multi-modal reasoning between point cloud and text data. Current 3D-VLMs directly embed the 3D point clouds into 3D tokens, following large 2D-VLMs with powerful reasoning capabilities. However, this framework has a great computational cost limiting its application, where we identify that the bottleneck lies in processing all 3D tokens in the Large Language Model (LLM) part. This raises the question: how can we reduce the computational overhead introduced by 3D tokens while preserving the integrity of their essential information? To address this question, we introduce Hierarchical Compensatory Compression (HCC-3D) to efficiently compress 3D tokens while maintaining critical detail retention. Specifically, we first propose a global structure compression (GSC), in which we design global queries to compress all 3D tokens into a few key tokens while keeping overall structural information. Then, to compensate for the information loss in GSC, we further propose an adaptive detail mining (ADM) module that selectively recompresses salient but under-attended features through complementary scoring. Extensive experiments demonstrate that HCC-3D not only achieves extreme compression ratios (approximately 98%) compared to previous 3D VLMs, but also achieves new state-of-the-art performance, showing the great improvements on both efficiency and performance.