Domain Generalization of Face Anti-Spoofing (DG-FAS) is increasingly becoming an important basis for model applications. Previous methods usually endeavor to project a domain-invariant feature space, which may not necessarily exist or be extremely difficult to explore. In this work, borrowing a solution from the large-scale vision-language models (VLMs) to replace direct editing of visual features, we propose a novel CLIP-Guided Semantic Alignment (CLIP-SA) for DG-FAS, which can guide visual feature alignment for generalization by leveraging the smoothness of the dense space of textual features. Specifically, in the textual branch, we design a Fine-Grained Prompt Template (FGPT) consisting of two language concepts, i.e., content and category, as supplements to visual features. Thus, the generalization of visual features can be improved by enhancing the model's understanding and association of content and category semantics. In the image encoder, we introduce a learnable Domain-Aware Adapter (DA-Adapter), which acts as a parameter-efficient transfer learning (PETL) method to transfer the pre-trained CLIP for the FAS task. Furthermore, the DA-Adapter contains an attention-based pooling layer that aggregates domain-invariant semantics as adaptive features for the output visual features. Finally, Our CLIP-SA demonstrates strong robustness to physical attacks on cross-domain benchmarks (ICMO, SCW), while also achieving excellent generalization against digital adversarial perturbations and on multiple Deepfake datasets.

Presentation Attack Detection (PAD) serves as a crucial safeguard for face recognition systems against presentation attacks such as printed photos, replayed videos, and 3D masks. Despite significant progress, existing PAD models still struggle to generalize across unseen domains due to variations in sensors, lighting, and attack materials. Recent Vision-Language Models (VLMs) have shown strong generalization ability, yet their applications in PAD remain limited because learned prompts, typically optimized under class-label supervision, fail to explicitly align with fine-grained attack-relevant visual semantics. As a result, the learned representations often overfit domain-specific artifacts instead of capturing transferable attack cues. To address this, we propose Concept-Informed Prompts Guided Presentation Attack Detection (CPG-PAD), a framework that introduces model-level concept guidance into the prompt learning process. Specifically, we design a Visual Concept-driven Enhancement (VCE) module that employs eXplainable AI (XAI) techniques to automatically discover PAD-relevant visual concepts and generate concept-associated heatmaps providing localized fine-grained guidance. Guided by these heatmaps, a Prompt-based Concept Injection (PCI) mechanism integrates these concepts into the prompt space through a Visual-Prompt Decoder (VPD) and a concept-mapping loss, enabling prompts to align with the model's internal concept space. This design enables CPG-PAD to capture generalizable and domain-invariant attack cues while effectively suppressing dataset-specific biases. Extensive experiments across nine benchmark datasets demonstrate that CPG-PAD consistently achieves state-of-the-art cross-domain performance under multi-source, limited-source, and single-source settings.

Face anti-spoofing, which aims to prevent the attacks of widely-used face recognition systems, is highly related to personal privacy and property security. However, existing benchmarks on face anti-spoofing mainly focus on RGB images, further challenged by consistently developed 3D high-fidelity (HiFi) masks. To facilitate the research on multimodal face anti-spoofing, we construct the HyperSpectral Face Anti-Spoofing (HySpeFAS) dataset. We introduce the newly-developed snapshot spectral imaging (SSI) technology to capture real and spoof faces, as well as identify unknown HiFi masks. Specifically, hyperspectral images (HSIs) acquired by SSI sensor contain rich information about the chemical composition of the targets, which can be used to effectively distinguish live human skin and various spoof materials. The HySpeFAS dataset contains 22,368 multimodal images (i.e., RGB, SSI, HSI) of 17 live subjects and 60 spoof subjects. Moreover, extensive experiments with baseline deep learning models validate the special features of the SSI images and the potential of SSI in FAS. By publishing the dataset as well as the baseline models, we encourage the community to foster the algorithm study associated with hyperspectral images and the development of SSI-equipped intelligent systems.

Recent advances in 3D human reconstruction have improved overall performance, yet current models still fail in the most challenging real-world scenarios. They often produce unstable geometry, inaccurate limb articulation and unreliable predictions under depth ambiguity or self-occlusion. A key reason is that existing datasets still lack the combination of high-resolution images, high-precision annotations and diverse whole-body motions required to support robust reconstruction. To address this gap, we present Human4K, a large-scale 4K multi-view whole-body human reconstruction dataset with mocap-accurate SMPL-X annotations. Human4K contains over six million 4K images captured by an eight-view high-resolution camera system synchronized with a professional Vicon motion capture setup, covering 11 subjects performing complex, highly articulated and strongly self-occluded full-body motions. All sequences are processed by a Motion-Retargeting and Refinement Module (MRRM) to ensure precise alignment for the full body and extremities. Experimental results show that training with Human4K consistently improves whole-body reconstruction on standard benchmarks, with particularly large gains for hands, feet and depth-ambiguous limb configurations.

Multimodal language models (LMs) have shown significant potential for applications across various domains but remain vulnerable to adversarial attacks. Current research in white-box or black-box settings generally struggles with unrealistic attack assumptions and limited efficacy of targeted attacks. This paper introduces CoGA, a novel gray-box collaborative adversarial attack method for multimodal LMs. Under our gray-box settings, attackers have access only to the victim model’s input encoders. With the guidance of different modalities, we perturb the embedding representations from encoders to disrupt the semantic alignment across modalities, ultimately causing inaccurate outputs on various downstream tasks. Specifically, we integrate text embeddings into the loss calculations of the image attack and utilize image embeddings to guide the ranking of vulnerable words and the selection of final samples. Extensive experiments demonstrate that our method achieves superior attack performance across diverse models and tasks, suggesting the shared vulnerability of multimodal LMs in confronting adversarial challenges. Our work provides new insights into the security of multimodal LMs, facilitating the deployment of more robust and secure models in practical applications.

Monteggia fractures exhibit a high missed diagnosis rate of over 20%, largely attributable to their subtle radiographic presentation and frequently low clinical suspicion. While AI-based diagnostic methods hold considerable potential to enhance detection accuracy, their development has been hampered by the absence of a dedicated, well-annotated imaging dataset. To address this gap, we introduce MFXR, a publicly available X-ray dataset designed to facilitate research in AI-based diagnosis of Monteggia fractures. The dataset comprises 4,586 X-ray images from 1,482 patients, including 2,793 Monteggia fracture images and 1,793 normal or other fracture controls. All cases have been annotated and validated by board-certified orthopedic surgeons, ensuring high-quality labels. Furthermore, we provide benchmark results based on eight representative deep learning models, offering performance baselines to support and guide future research. This public dataset and its accompanying benchmarks are expected to accelerate the development of reliable AI diagnostic tools, thereby helping to reduce diagnostic errors and improve patient care in the management of Monteggia fractures.

The challenge of Domain Generalization (DG) in Face Anti-Spoofing (FAS) is the significant interference of domain-specific signals on subtle spoofing clues. Recently, some CLIP-based algorithms have been developed to alleviate this interference by adjusting the weights of visual classifiers. However, our analysis of this class-wise prompt engineering suffers from two shortcomings for DG FAS: (1) The categories of facial categories, such as real or spoof, have no semantics for the CLIP model, making it difficult to learn accurate category descriptions. (2) A single form of prompt cannot portray the various types of spoofing. In this work, instead of class-wise prompts, we propose a novel Content-aware Composite Prompt Engineering (CCPE) that generates instance-wise composite prompts, including both fixed template and learnable prompts. Specifically, our CCPE constructs content-aware prompts from two branches: (1) Inherent content prompt explicitly benefits from abundant transferred knowledge from the instruction-based Large Language Model (LLM). (2) Learnable content prompts implicitly extract the most informative visual content via Q-Former. Moreover, we design a Cross-Modal Guidance Module (CGM) that dynamically adjusts unimodal features for fusion to achieve better generalized FAS. Finally, our CCPE has been validated for its effectiveness in multiple cross-domain experiments and achieves state-of-the-art (SOTA) results.

Face recognition remains vulnerable to presentation attacks, calling for robust face anti-spoofing (FAS) solutions. Recent Multimodal Large Language Model (MLLM)-based FAS methods reformulate the binary classification task as the generation of brief textual descriptions to improve cross-domain generalization. However, their generalizability is still limited, as such descriptions mainly capture intuitive semantic cues (e.g., screen borders or mask contours) while struggling to perceive fine-grained visual patterns. To address this limitation, we incorporate external visual tools into MLLMs to encourage deeper investigation of subtle spoof clues. Specifically, we propose the Tool-Augmented Reasoning FAS (TAR-FAS) framework, which reformulates the FAS task as a Chain-of-Thought with Visual Tools (CoT-VT) paradigm, allowing MLLMs to begin with intuitive observations and adaptively invoke external visual tools for fine-grained investigation. To this end, we design a tool-augmented data annotation pipeline and construct the ToolFAS-16K dataset, which contains multi-turn tool-use reasoning trajectories. Furthermore, we introduce a tool-aware FAS training pipeline, where Diverse-Tool Group Relative Policy Optimization (DT-GRPO) enables the model to autonomously learn efficient tool use. Extensive experiments under a challenging one-to-eleven cross-domain protocol demonstrate that TAR-FAS achieves state-of-the-art performance while providing fine-grained visual investigation for trustworthy spoof detection.

Scene-level point cloud self-supervised learning (PC-SSL) has demonstrated potential in enhancing the generalization capability of 3D vision models. Despite the advances achieved in the field through existing methods, the sample-independent modelling paradigm still poses significant limitations in terms of maintaining consistent semantic representations across different scenes. This challenge hinders the construction of a unified and transferable semantic space. To address this issue, we propose a PC-SSL framework based on cross-sample semantic propagation (CSP), in which samples within a batch are serialized into a continuous input and processed by a state-space model to enable semantic state propagation. This mechanism explicitly models the dynamic dependencies across samples in the state space, allowing the network to establish cross-sample semantic consistency in the latent space, and thereby achieve global semantic alignment. Since serialization-based pretraining requires batch-level input organization, we further introduce an asymmetric semantic preservation distillation (SPD) during finetuning to achieve structural alignment of semantic transfer and eliminate inconsistencies caused by batch dependency. The proposed SPD ensures stable transfer of pretrained semantics through a heterogeneous input mechanism and a semantic feature alignment constraint. This enables the model to maintain structured semantic consistency and robustness under single-scene testing conditions. Extensive experiments on multiple benchmark datasets demonstrate that our method consistently outperforms state-of-the-art methods in both performance and semantic consistency.

Flexible Modal Face Anti-Spoofing (FAS) aims to improve the model’s adaptability to any given deployment modality. The current methods commonly use multiple independent encoders to learn modality-general features for flexible testing, which inevitably involves redundant computation. In this work, we propose a novel Flexible Modal Mixture-of-Experts (FM-MoE) for adaptively learning the representation of multi-modal data in a unified framework, which includes Modal Inter-MoE and Modal Intra-MoE. Specifically, considering that the challenge of flexible modal tasks lies in significant modal differences that make it difficult for classifiers to identify subtle spoofing clues, we employed Modal Inter-MoE to learn modality-general and modality-specific features, respectively. Meanwhile, to more effectively capture multi-modal visual information, the Modal Intra-MoE replaces the traditional routing patch token strategy, focusing on extracting instance-level spoofing clues from the channel perspective in the feature map. In addition, to avoid the imbalance problem caused by simultaneous training of multi-modal data, we introduce Modal Inter Knowledge Distillation Loss (MI-KDL), which includes Cross-Modal Loss and Shared to Multi-Modal Loss (CML & S2MML). CML is designed to enhance information alignment across different modalities, while S2MML endeavors to transfer shared knowledge to all modalities, preventing the model from excessively optimizing certain single modalities. We conduct extensive experiments to demonstrate the flexibility and superiority of our FM-MoE framework over the state-of-the-art methods.

Domain generalization (DG) based Face Anti-Spoofing (FAS) aims to improve the model’s performance on unseen domains. Existing methods either rely on domain labels to align domain-invariant feature spaces or disentangle generalizable features from the whole sample, which inevitably leads to the distortion of semantic structures and achieves limited generalization. In this work, we make use of large-scale VLMs like CLIP and leverage the textual features to dynamically adjust the classifier’s weights for exploring generalizable visual features. Specifically, we propose a novel CLIP with Instance and Category Prompts Engineering for DG FAS (ICPE-FAS), which includes two text branches: (1) Instance-wise Content Prompt Learning (ICPL) to generate instance-prompts consisting of two lightweight transformers, namely Content Q-Former (CQF) and Style Q-Former (SQF), to learn the different semantic prompts conditioned on content and style features by using a set of learnable query vectors, respectively. The generalizable content prompt can be learned by two innovations: (a) A Prompt-Text Matched (PTM) supervision is introduced to ensure CQF learns a visual representation that is most informative of the content description. (b) A Diversified Style Prompt (DSP) technology is proposed to diversify the learning of style prompts by mixing feature statistics between instance-specific styles. Finally, the learned text features modulate visual features to generalization through the designed Prompt Modulation (PM). (2) Category-wise Segmented Prompt Learning (CSPL) to generate Domain-agnostic Prompt (DaP) and Domain-specific Prompt (DsP), which can serve as weights for the classifier to accurately suppress domain-related signals. Extensive experiments show that the ICPE-FAS is effective and outperforms the state-of-the-art methods on several cross-domain benchmarks.

Face antispoofing (FAS) technologies play a pivotal role in safeguarding face recognition (FR) systems against potential security loopholes. The biometric community has witnessed significant advancements lately, largely due to the exceptional performance of deep learning architectures and the abundance of substantial datasets. Despite these progress, FR systems remain susceptible to both physical and digital forgery attacks. However, most existing detection methods do not address both types of threats concurrently. To bridge this gap and foster the development of a comprehensive detection framework, we have compiled a unified dataset named UniAttackData. This dataset incorporates both physical and digital spoofing attacks while maintaining identity consistency, encompassing 1800 participants each subjected to two different physical attacks (PAs) and 12 different digital attacks (DAs), respectively. This effort has resulted in a comprehensive collection of 29,706 video samples. We organized the Chalearn FAS face attack detection challenge based on this novel resource to boost research aiming to promote joint antispoofing efforts. The Chalearn unified antispoofing attack detection challenge drew 136 teams during the development phase, with 13 teams advancing to the final round. The organizing team revalidated and re-executed the submitted code to determine the final rankings. This paper provides a summary of the challenge, covering the dataset used, the protocol definition, the evaluation metrics, and the competition results. Additionally, we discuss the top-ranked algorithms and the research insights offered by the competition for attack detection.

Human parsing, which classifies each pixel in human images into predefined classes of human parts or clothing, faces challenges in efficient modeling complex scenarios involving pose variations and occlusion. While existing methods have shown that pose estimation benefits human parsing via multi-task learning, most of them use simplistic feature fusion strategies, making it difficult to achieve efficient parsing-pose interaction. In this paper, we propose Mamba-Driven Feature Fusion Network (MFNet), leveraging Mamba’s global modeling capabilities to establish deep bidirectional parsing-pose interaction, enhancing fused features for joint human parsing and pose estimation. MFNet employs a shared backbone for initial feature extraction. Central to the framework, we introduce a cross fusion mamba module, enabling efficient parsing-pose interaction while capturing comprehensive global information. Simultaneously, an attention atrous spatial pyramid pooling module refines parsing features using dual channel-spatial attention to adaptively enhance multi-scale context. Through the combination of the two modules, MFNet effectively integrates parsing and pose features while capturing comprehensive global information. Results on two datasets demonstrate MFNet’s superiority in various metrics.

Point cloud tasks have recently benefited from Mamba-based architecture, which leverage state space modeling to achieve strong performance. Previous studies have primarily focused on network design while overlooking the importance of position encoding and relying on coarse-grained geometric feature aggregation. The former leads to semantic ambiguity due to inconsistent spatial relationships, while the latter results in geometric feature dispersion by overlooking fine-grained local geometric details. To tackle the above problem, we propose a novel framework, PointMC, including Multi-view Consistent Learnable Position Encoding (MCLPE) and Center-Global Feature Fusion (CGFF), to provide semantically coherent positional guidance for inter-patch and enable fine-grained geometric structure aggregation within intra-patch regions. Specifically, the proposed MCLPE module is inspired by a spatial structure modeling mechanism guided by physical constraints, leverages multi-view virtual reconstruction and a learnable strategy to dynamically constrain spatial relationships along patch boundaries, thereby enhancing the semantic consistency and representational clarity across inter-patch regions. Furthermore, considering the lack of local structural information within each patch, the CGFF module employs a dual-guidance mechanism based on center and global structures to effectively promote the aggregation of local geometric features. Extensive experiments on multiple benchmark datasets validate the effectiveness of PointMC, consistently outperforming existing state-of-the-art methods, and demonstrating superior capability in capturing both inter-patch semantic consistency and intra-patch geometric details.

The overfitting of domain signals results in poor domain generalization of face anti-spoofing. The current methods usually improve the diversity of source domains to alleviate this overfitting. However, this benefit is minimal, as even the most diverse domain signals will also be absent in the target domain. In this work, we propose a Domain-Guided Prompt Distribution Learning (DGPDL) built on Vision-Language Models like CLIP, which explores a unified representation of domain signals as a prompt across the source and target domain to alleviate the understanding bias caused by domain gaps. Specifically, we first define a learnable Domain-Specific Distribution (DSD) that covers as many domain elements as possible, such as image quality, color tone, camera settings, etc., which establish connections between different domains and linearly combinable prompt in any domain; Then, based on the style statistics of the given sample, we construct its optimal Domain-Specific Prompts (DSPs) from the defined DSD through the designed Prompt Assemble Attention (PAA) with the similarity matching; Finally, the assembled DSPs will act as carrier or agent to perform on both the vision and language branches, synergistically improving the model's recognition of domain signals. By using the prompt to represent domain signals uniformly, if the model can be robust to DSPs in the source domain, it should be applicable to target domain, as they share the same DSD. By representing domain signals as prompts rather than instantiation features, DGPDL effectively reduces the reliance on specific domain appearances. This design enables the model to dynamically adapt to unseen target domains without the need for retraining. Extensive experiments show that the DGPDL is effective and outperforms the state-of-the-art methods on several cross-domain benchmarks.

Monocular and generative services can now predict visually convincing per-frame depth for everyday videos. Yet these predictions are scale-ambiguous, temporally unstable, and pose-free, which prevents reliable use in mapping, AR re-localization, and 3D reconstruction. We present Geo-Aligner, a geometry-centered pipeline that upgrades arbitrary per-frame depth into scale-consistent video geometry together with camera trajectories. Geo-Aligner first aggregates cross-frame evidence to build a pose-free geometric scaffold, which provides reliable multi-view cues without requiring ground-truth poses. To resolve the inherent scale ambiguity and improve local structure, we introduce a Scale-Consistent Depth Refiner (SCDR) that fuses semantic tokens with geometric tokens using temporal positional encoding and cross-attention. SCDR injects global semantics to guide local geometry, sharpens boundaries, suppresses flicker, and enforces a shared scale across frames. In parallel, we incorporate monocular surface priors as soft anchors inside the decoder pathway, stabilizing low-texture or occluded regions without over-constraining the solution. Across diverse scenes and dynamic sequences, Geo-Aligner produces metric depths and consistent trajectories that are more temporally stable and less failure-prone than representative alternatives. The resulting outputs plug directly into MVS, NeRF/3DGS, and AR pipelines, turning per-frame depth into actionable, video-level geometry.

Knowledge distillation has proven highly effective in anomaly detection by quantifying discrepancies between teacher and student (T-S) networks. However, existing approaches predominantly focus on local feature distillation, often neglecting the contextual relationships among features. This oversight leads to two main issues: (1) the student network captures only fragmented local information and fails to learn the distribution of normal data thoroughly, and (2) the lack of contextual information hinders accurate anomaly localization, as anomalies typically manifest as inconsistencies between local features and their global context. To address these issues, we propose a simple yet effective Context-aware Knowledge Distillation (ContextKD) framework for anomaly detection. Specifically, we integrate a Context-Aware Reasoning (CAR) block after each layer of the student network to enhance contextual learning. The CAR block consists of two key modules: the Layer-wise Context Mining (LCM) module and the Context-Guided Reconstruction (CGR) module. The LCM module extracts rich contextual information from multi-layer features via cross-layer interactions, while the CGR module randomly masks local feature tokens and leverages contextual relationships to reconstruct the masked regions. This design enables the student network to capture the distribution of normal data and amplify feature inconsistencies within anomalous regions. Furthermore, we systematically investigate 16 distinct T-S distillation frameworks and compare their anomaly detection performance under standardized settings. To validate the effectiveness of ContextKD, we conducted extensive experiments across 30 diverse categories collected under varying imaging conditions (e.g., lighting, resolution, and background noise). Additionally, a real-world oil seal inspection application further demonstrates the effectiveness and practicality of the proposed ContextKD, achieving a 98.5% detection accuracy with less than 2% false alarms under production-line standards.

Multi-modal face anti-spoofing (FAS) is receiving increasing attention due to the complementarity between different modal sample information. However, its performance is limited by the sparse availability of multi-modal data (i.e., Depth, NIR) compared to visible light data (i.e., RGB). In this paper, we intend to utilize a large amount of RGB data to improve the performance of FAS on Depth or NIR data with only few-shot target modal samples, which is essentially a cross-modal generalization FAS (CMG-FAS) task. The difficulty of analyzing the CMG-FAS lies in the classifier is seriously interfered with by modal-gap due to large distribution discrepancies among different modalities. To address this problem, we introduce a novel cross-modal contrastive language-image pre-training framework termed CM-CLIP, that aims to leverage the textual feature to dynamically adjust the classifier’s weights for exploring generalizable visual features. Specifically, in the text branch of CM-CLIP, a fine-grained prompt learning (FGPL) strategy is introduced to describe a sample from three levels: content, modality, and category. Furthermore, our prompts are designed to be modally generalized by two contributions: 1) Reducing the correlation between content and modality prompts. 2) Combining the learnable context vectors with an input-conditional token generated for each image. Additionally, we design an adaptive differential convolution adapter (ADC-Adapter) into the image encoder of CM-CLIP, which can effectively reduce the number of learnable parameters and bridge the task-gap between general image recognition and FAS task quickly. Extensive experiments show that the CM-CLIP framework is effective and outperforms the state-of-the-art methods on several multi-modal benchmarks.

Optical Coherence Tomography (OCT) is a core imaging technology for detecting pathological biomarkers associated with Diabetic Retinopathy (DR) and Diabetic Macular Edema (DME).Leveraging its high-resolution advantage, it can clearly visualize retinal microstructures and biomarker characteristics. Current mainstream OCT biomarker detection methods are mostly confined to single-modality imaging analysis, failing to fully exploit the complementary multimodal information such as clinical data. This results in inadequate capture of complex pathological features of diseases by models. This paper proposes a novel multimodal active learning framework driven by reinforcement learning (MALRL) for DR/DME-related OCT biomarker classification.We design an Actor-Critic agent to dynamically adjust modality importance weights, achieving adaptive modality balancing in the sample selection process.Moreover, in DR/DME-related OCT biomarker classification, different biomarkers exhibit heterogeneous feature distributions, including variations in visual granularity, class imbalance, and modality dependence. To address this challenge, we employ a feedback-driven mechanism that combines intrinsic rewards and extrinsic rewards: the intrinsic reward evaluates candidate samples according to informativeness, representativeness, and modality balance, while the extrinsic reward reinforces sampling decisions that reduce validation loss. This enables the RL agent to adaptively adjust modality weights and select informative yet representative samples for biomarker categories.Experimental results on the public OLIVES dataset demonstrate that the MALRL framework can integrate complementary information from OCT images and clinical indicators (e.g., best-corrected visual acuity (BCVA), central subfield thickness (CST)), improving classification performance and generalization of multimodal models.