Visual decoding from electroencephalography (EEG) has emerged as a highly promising avenue for non-invasive brain-computer interfaces (BCIs). Existing EEG-based decoding methods predominantly align brain signals with the final-layer semantic embeddings of deep visual models. However, relying on these highly abstracted embeddings inevitably leads to severe cross-modal information mismatch. In this work, we introduce the concept of Neural Visibility and accordingly propose the EEG-Visible Layer Selection Strategy, aligning EEG signals with intermediate visual layers to minimize this mismatch. Furthermore, to accommodate the multi-stage nature of human visual processing, we propose a novel Hierarchically Complementary Fusion (HCF) framework that jointly integrates visual representations from different hierarchical levels. Extensive experiments demonstrate that our method achieves state-of-the-art performance, reaching an 84.6

Recently, instruction-based image editing has made significant progress. However, human-centric editing remains challenging due to the difficulty of preserving facial and body feature consistency. Moreover, most existing methods handle either rigid edits (e.g., object addition or replacement) or non-rigid edits (e.g., viewpoint changes), which limits their ability to perform diverse editing types. To address these issues, we introduce HumanEdit, the first human-centric editing dataset, comprising both non-rigid (expression and pose changes) and rigid (clothing replacement) editing data. Additionally, we propose a MoE-Unet denoising architecture that consists of two denoising expert networks, each of them specialized in one editing type, and a gating network named GateNet that can dynamically route each instruction to the corresponding expert. Extensive experiments on the HumanEditEval benchmark show that our method achieves state-of-the-art performance on both types of human-centric editing, generating superior and more natural edited images.

As one of the most established biometric modalities, fingerprint recognition plays a crucial role in personal authentication and security systems, with fingerprints captured through contact-based or contactless acquisition. However, contactless fingerprint data remains relatively scarce compared to contact-based data, posing challenges for both intra-domain and cross-domain matching. To address these issues, we propose a contact-to-contactless fingerprint generation approach, which integrates a GAN-based fingerprint transfer module with a Unified Fingerprint Enhancement (UFE) module to maintain content consistency and preserve identity information. Extensive experiments on the NIST SD14 database, including both quantitative and qualitative evaluations, demonstrate the effectiveness of the proposed method.

Contactless fingerprint has gained lots of attention in recent fingerprint studies. However, most existing contactless fingerprint algorithms treat contactless fingerprints as 2D plain fingerprints, which lack consideration of the modality difference between contactless and contact fingerprints, especially the intrinsic 3D features in contactless fingerprints. This paper proposes a novel contactless fingerprint recognition algorithm that captures the revealed 3D feature of contactless fingerprints rather than the plain 2D feature. The proposed method first recovers 3D features from the monocular input contactless fingerprint, including the 3D shape model and 3D fingerprint feature (minutiae, orientation, etc.). Then, a novel pose estimation and matching method based on 3D graph network is proposed according to the extracted 3D feature. The proposed method is able to perform robust 3D feature extractions on various contactless fingerprints across multiple finger poses and correcting the pose of contactless fingerprints in 3D space. The results of the experiments on contactless fingerprint databases show that the proposed method successfully improves the matching accuracy of contactless fingerprints. Exceptionally, our method performs stably across multiple poses of contactless fingerprints due to 3D pose regression and embeddings, which is a great advantage compared to 2D-based previous contactless fingerprint recognition algorithms.

Text-based person retrieval (TBPR) aims to retrieve pedestrian images using natural language descriptions and plays an important role in public welfare such as "searching for missing children". Despite recent advances, existing CLIP-based methods are computationally expensive, making them unsuitable for resource-constrained scenarios. To address this problem, we propose a novel Relational Dual-granularity Distillation framework (RDD) for TBPR, which consists of three key components: a Clarity-aware Data Augmentation (CDA) module is first designed to improve the diversity of training samples, a Dual-granularity Distillation Alignment framework is then presented to enhance knowledge transfer at both global and local levels, and finally the Memory-guided Relational Learning (MRL) is proposed to capture structual patterns in the teacher’s internal representations. Experiments on three widely used benchmarks demonstrate that our distillation framework significantly improves upon the TinyCLIP baseline, obtaining notable performance gains with the R@1, R@5 and mAP increasing by 5.38%, 3.49%, and 4.17%, respectively. Furthermore, it demonstrates a reduction in parameter size, FLOPs, and inference latency compared to CLIP-based models, suggesting its potential for more efficient real-world applications. The code is available at https://github.com/JThuge/RDD.

We present an empirical study of embedding-based retrieval under realistic conversational settings, where queries are short, dialogue-like, and weakly specified, and retrieval corpora contain structured conversational artifacts. Focusing on Qwen3-embedding models, we identify a deployment-relevant robustness vulnerability: under conversational retrieval without query prompting, structured dialogue-style noise can become disproportionately retrievable and intrude into top-ranked results, despite being semantically uninformative. This failure mode emerges consistently across model scales, remains largely invisible under standard clean-query benchmarks, and is significantly more pronounced in Qwen3 than in earlier Qwen variants and other widely used dense retrieval baselines. We further show that lightweight query prompting qualitatively alters retrieval behavior, effectively suppressing noise intrusion and restoring ranking stability. Our findings highlight an underexplored robustness risk in conversational retrieval and underscore the importance of evaluation protocols that reflect the complexities of deployed systems.

Spatial transcriptomics (ST) enables spatially resolved gene profiling but remains expensive and low-throughput, limiting large-cohort studies and routine clinical use. Predicting spatial gene expression from routine hematoxylin and eosin (H E) slides is a promising alternative, yet under realistic leave-one-slide-out evaluation, existing models often suffer from slide-level appearance shifts and regression-driven over-smoothing that suppress biologically meaningful variation. CHRep is a two-phase framework for robust histology-to-expression prediction. In the training phase, CHRep learns a structure-aware representation by jointly optimizing correlation-aware regression, symmetric image-expression alignment, and coordinate-induced spatial topology regularization. In the inference phase, cross-slide robustness is improved without backbone fine-tuning through a lightweight calibration module trained on the training slides, which combines a non-parametric estimate from a training gallery with a magnitude-regularized correction module. Unlike prior embedding-alignment or retrieval-based transfer methods that rely on a single prediction route, CHRep couples topology-preserving representation learning with post-hoc calibration, enabling stable neighborhood retrieval and controlled bias correction under slide-level shifts. Across the three cohorts, CHRep consistently improves gene-wise correlation under leave-one-slide-out evaluation, with the largest gains observed on Alex+10x. Relative to HAGE, the Pearson correlation coefficient on all considered genes [PCC(ACG)] increases by 4.0

Reconstructing external visual stimuli is vital for understanding human cognition. The existing video reconstruction methods usually map brain activity to corresponding stimuli responses through multi-modal contrastive learning. However, the neural representations obtained by these methods are not always effective, mainly for two reasons. Firstly, the data of functional magnetic resonance imaging (fMRI)-video pairs is extremely scarce. Secondly, the heterogeneity between modalities are neglected. To address these challenges, we introduce a novel two-stage brain decoding framework, Mind2Word, for high-quality video reconstruction. In the first stage, a Mind-Centric (MC) encoder is firstly proposed to map fMRI signals into pseudo-words, which are further refined within the text embedding space to effectively obtain generalized representations of videos corresponding to fMRI data while bridging the modality gap. In the second stage, guided by the pseudo-words, a video generator is introduced to reconstruct high-quality videos. Additionally, by integrating pseudo-words with style prompts, Mind2Word enables the flexible creation of artistic and stylized videos. Mind2Word achieves state-of-the-art (SOTA) performance across three benchmarks, with a 38% improvement in structural similarity index (SSIM). Furthermore, it breaks through the limitation of the per-subject-per-model paradigm and achieves superior performance in cross-subject video reconstruction with just a single model, demonstrating a strong generalization capability.

Precise and controllable image editing, especially object removal and insertion, represents one of the most common demands in image manipulation. However, existing methods suffer from severe limitations. Mask-based inpainting often introduces visual artifacts and semantic inconsistencies, while instruction-based approaches lack accurate spatial control and tend to unintentionally modify background regions. To address these issues, we propose two key contributions. First, we develop a fully automated and self-improving pipeline for synthetic data generation. This pipeline utilizes a Large Language Model (LLM) to generate diverse prompts, a Diffusion Transformer (DiT) fine-tuned evolutionarily to synthesize high-quality images, and a Multimodal LLM (MLLM) combined with open-set object detector for automated quality control and annotation. This process produces the Remove/Add Dataset (RAD), consisting of over 514,510 high-quality image pairs, each richly annotated with bounding boxes, segmentation masks, and a variety of editing instructions. Second, based on RAD, we introduce Remove/Add Anything (RAA), a novel editing framework with precise spatial control. Built upon a diffusion-based inpainting model, RAA achieves high editing accuracy by conditioning on both textual instructions and an explicitly defined region of interest (ROI), enabling efficient fine-tuning while maintaining global visual coherence. Extensive experiments demonstrate that RAA significantly outperforms existing open-source methods on both addition and removal tasks, and even slightly surpasses costly proprietary models.

Diffusion Transformer(DiT)-based video generation models exhibit strong generative capabilities. Trajectory control represents a user-friendly task in the field of controllable video generation. However, existing methods either require substantial training resources or are specifically designed for U-Net, do not take advantage of the superior performance of DiT. To address these issues, we propose a simple but effective training-free framework for trajectory control in text-to-video generation with decoupled guidance, tailored for DiT. Specifically, first, we propose foreground-background separation guidance: we use the Large Language Model (LLM) to convert user-provided prompts into foreground and background prompts, which respectively guide the generation of foreground and background regions in the video. Then, we analyze 3D full attention and explore the tight correlation between inter-token attention scores and position embedding. Based on this, we propose inter-frame Spatial-Temporal Decoupled 3D-RoPE (STD-RoPE). By modifying only foreground tokens' position embedding, STD-RoPE eliminates their cross-frame spatial discrepancies, strengthening cross-frame attention among them and thus enhancing trajectory control. Additionally, we achieve 2.5D-aware trajectory control by regulating the density of position embedding. Extensive experiments demonstrate that our method outperforms previous methods in both video quality and trajectory controllability.

Deep learning algorithms have driven substantial progress in remote sensing semantic segmentation. However, conventional approaches typically rely on predefined semantic categories, necessitating costly data annotation and model retraining when new classes are introduced. While large-scale vision-language models (VLMs), such as contrastive language image pre-training (CLIP), enable segmentation of arbitrary class with natural language guidance, their limited localization capability poses challenges for dense prediction tasks. This study investigates the potential of CLIP for semantic segmentation of optical remote sensing images (RSIs) and proposes LDSeg, a novel training-free framework that optimizes localization ability and visual-textual alignment through dual representation refinement. Through systematic analysis, we identify anomalous "singularity" tokens in the CLIP visual encoder that aggregate global contextual information while disproportionately attracting attention from local patch tokens, thereby degrading spatial discriminability. To address this, we introduce a singularity feature repair (SFR) strategy that mitigates feature distortion by recalibrating these tokens. Furthermore, we develop a hierarchical semantic expansion (HSE) method to generate precise hierarchical text descriptions, enhancing cross-modal alignment. The SFR and HSE strategies complement existing methods, providing further improvements. Extensive experiments demonstrate that LDSeg achieves state-of-the-art performance, delivering mean intersection over union (mIoU) improvements of 0.95%-7.84% over existing methods.

Image Aesthetic Assessment (IAA) aims to quantify the aesthetic quality of images, while Image Aesthetic Captioning (IAC) describes the aesthetic characteristics of images. For the IAA task, existing methods often rely solely on human-annotated scores to fine-tune pretrained visual models, making it difficult to perceive the impact of different attributes such as style, semantic content, and scene. Moreover, IAA performance is further hampered by the inherent constraints of single-modal aesthetic capabilities. Meanwhile, IAC quality is also restricted by noisy data and model capacity. To address these issues, we propose a Multi-Attribute Aesthetic Assessment and Captioning (MA3C) framework. First, we pretrain a vision-language model that leverages various image-attribute pairs and aesthetic image-comment pairs for training, thereby endowing the model with the ability to generate aesthetic captions and extract multiple attributes and multimodal aesthetic features. Second, to improve IAC quality, we propose a Caption-Filter method to purify noisy aesthetic comment datasets and introduce a pre-trained Large Language Model (LLM) to enhance caption generation. Finally, in the IAA task, comprehensive multimodal aesthetic features are learned by fusing style, semantic content, and scene attributes. Experimental results demonstrate that our model achieves superior performance on the AVA-Captions and AVA datasets, obtains competitive results on the AADB, TAD66K, and PARA datasets, and exhibits promising zero-shot IAA capability.

Recently, text-to-video generation has developed rapidly due to the crucial role of video diffusion models (VDMs). Although the customization of text-to-video generation models has gained growing attention, customizing motion dynamics still poses a significant challenge. The existing methods are either limited by their inability to decouple appearance information or are prone to be affected by background variations. To address these issues, we propose a novel one-shot tuning framework, enabling it to generate videos that maintain motion consistency with a single reference video across different contexts. Specifically, our method focuses on the temporal layers in VDM and introduces a Spatial-Attention-Guided motion learning objective (SAG), which first leverages spatial attention maps extracted from the spatial transformer blocks in VDM to construct Motion-Mask that highlights foreground dynamics. Secondly, the Motion-Mask is employed to constrain the learning of motion patterns, driving the model to focus on semantically meaningful dynamics while suppressing background variations. Extensive experiments on diverse datasets demonstrate that our approach outperforms multiple state-of-the-art methods in both quantitative and qualitative evaluations.

Data pruning is critical for adapting large-scale foundation models to downstream tasks under constrained computational and data resources. However, many existing methods rely on shallow statistics that ignore the semantic relationships between images and text, which is particularly harmful in multimodal settings. We propose multi-VLM Collaborated Adaptive Sampling (mCAS), a data-pruning framework that exploits the complementary strengths of multiple pre-trained vision–language models (VLMs). mCAS comprises three components: a semantic enhancement module that uses large language models to refine textual descriptions and enrich image representations; a multi-VLM collaboration module that aggregates ordinal rankings from diverse VLMs through a training-free, equal-weight consensus strategy; and an adaptive sampling module that selects informative samples from a mid-range similarity spectrum, reducing redundancy while mitigating overfitting and underfitting. Experiments on Mini-ImageNet, COCO-GLT, Caltech101, and ImageNet-1K demonstrate that mCAS consistently outperforms representative pruning baselines across different few-shot and transfer settings. On Mini-ImageNet, mCAS improves 5-shot and 10-shot classification accuracy by up to 41.3 and 32.6 percentage points, respectively, while using fewer training samples. On ImageNet-1K, it further improves classification accuracy by 8.24, 9.72, and 8.81 percentage points under the 5-shot, 10-shot, and 15-shot settings, respectively. Additional ablation studies verify the contribution of each component and confirm that multi-VLM collaboration yields more stable and reliable pruning decisions. These results show that leveraging multi-VLM collaboration and adaptive sampling provides an effective way to preserve critical semantic information, enabling more reliable and efficient deployment of foundation models in data-scarce multimodal environments.

Knowledge editing facilitates precise and targeted updates in Large Language Models (LLMs) and Multimodal Models (LMMs) without the need for full retraining. Although existing editing methods achieve remarkable performance in textual modality, they still struggle to simultaneously preserve pre-trained knowledge and generalize new knowledge in intricate multimodal contexts. To address this, we propose KSpace, a novel solution that decomposes the update into two complementary subspaces: (1) a null space aligned with preserved knowledge, constructed via multimodal initialization to maintain the model’s general capabilities, and (2) a knowledge space extracted from the model’s internal states of multimodal perturbations, capturing invariant semantics of updates. Extensive experiments on various LMMs, including LLaVA-v1.5-7B, Qwen2.5-VL-7B, and Phi-4-multimodal, show that KSpace outperforms LoRA-based methods by an average of 14.17% accuracy across three metrics: reliability, generality, and locality under rigorous LLM-as-a-Judge evaluation, which demonstrates that KSpace can achieve superior real-world editing quality.

Zero-shot Composed Image Retrieval (ZS-CIR) aims to retrieve the target image based on a reference image and a text description without requiring in-distribution triplets for training. One prevalent approach follows the vision-language pretraining paradigm that employs a mapping network to transfer the image embedding to a pseudo-word token in the text embedding space. However, this approach tends to impede network generalization due to modality discrepancy and distribution shift between training and inference. To this end, we propose a Data-efficient Generalization (DeG) framework, including two novel designs, namely, Textual Supplement (TS) module and Semantic Sample Pool (SSP) module. The TS module exploits compositional textual semantics during training, enhancing the pseudo-word token with more linguistic semantics and thus mitigating the modality discrepancy effectively. The SSP module exploits the zero-shot capability of pretrained Vision-Language Models (VLMs), alleviating the distribution shift and mitigating the overfitting issue from the redundancy of the largescale image-text data. Extensive experiments over four ZS-CIR benchmarks show that DeG outperforms the stateof-the-art (SOTA) methods with much less training data, and saves substantial training and inference time for practical usage.

Referring Multi-Object Tracking (RMOT) aims to track multiple objects specified by natural language expressions in videos. With the recent significant progress of one-stage methods, the two-stage Referring-by-Tracking (RBT) paradigm has gradually lost its popularity. However, its lower training cost and flexible incremental deployment remain irreplaceable. Rethinking existing two-stage RBT frameworks, we identify two fundamental limitations: the overly heuristic feature construction and fragile correspondence modeling. To address these issues, we propose FlexHook, a novel two-stage RBT framework. In FlexHook, the proposed Conditioning Hook (C-Hook) redefines the feature construction by a sampling-based strategy and language-conditioned cue injection. Then, we introduce a Pairwise Correspondence Decoder (PCD) that replaces CLIP-based similarity matching with active correspondence modeling, yielding a more flexible and robust strategy. Extensive experiments on multiple benchmarks (Refer-KITTI/v2, Refer-Dance, and LaMOT) demonstrate that FlexHook becomes the first two-stage RBT approach to comprehensively outperform current state-of-the-art methods. Code can be found in the Supplementary Materials.

As a challenging vision-language task, Zero-Shot Composed Image Retrieval (ZS-CIR) is designed to retrieve target images using bi-modal (image+text) queries. Typical ZS-CIR methods employ an inversion network to generate pseudo-word tokens that effectively represent the input semantics. However, the inversion-based methods suffer from two inherent issues: First, the task discrepancy exists because inversion training and CIR inference involve different objectives. Second, the modality discrepancy arises from the input feature distribution mismatch between training and inference. To this end, we propose a lightweight post-hoc framework, consisting of two components: (1) A new text-anchored triplet construction pipeline leverages a large language model (LLM) to transform a standard image-text dataset into a triplet dataset, where a textual description serves as the target of each triplet. (2) The MoTa-Adapter, a novel parameter-efficient fine-tuning method, adapts the dual encoder to the CIR task using our constructed triplet data. Specifically, on the text side, multiple sets of learnable task prompts are integrated via a Mixture-of-Experts (MoE) layer to capture task-specific priors and handle different types of modifications. On the image side, MoTa-Adapter modulates the inversion network's input to better match the downstream text encoder. In addition, an entropy-based optimization strategy is proposed to assign greater weight to challenging samples, thus improving adaptation efficiency. Experiments show that, with the incorporation of our proposed components, inversion-based methods achieve significant improvements, reaching state-of-the-art performance across four widely-used benchmarks.