Collaborative perception extends the perceptual range of autonomous vehicles by sharing information across agents, but heterogeneous sensors and perception models make intermediate feature fusion difficult to deploy at scale. Existing heterogeneous collaboration methods typically follow a translation-first paradigm: collaborator features must be aligned, adapted, or projected into an ego-compatible space before fusion. Such feature-compatibility contracts improve fixed-system performance, but they couple deployment to collaborator-specific adaptation and make newly joined heterogeneous agents costly to integrate. To address this gap, we propose INTACT, an ego-guided typed sparse evidence retrieval framework for heterogeneous collaborative perception. Instead of translating an entire collaborator feature map, INTACT lets the ego vehicle issue typed evidence queries that express suspected objects and evidence-deficient regions. Collaborators respond only with local evidence at queried locations, and the ego selects useful responses through sparse per-query routing and injects them through gated residual write-back. This changes the compatibility requirement from global feature-map interpretability to local, typed response comparability under ego-issued queries, enabling a zero-training heterogeneous insertion protocol in which the ego interface is trained once and new collaborators join through checkpoint merging. Extensive experiments on simulated and real-world heterogeneous collaborative perception benchmarks validate the effectiveness and deployability of INTACT. On OPV2V-H, INTACT achieves 80.1 AP70 with only 0.52M additional parameters and 18.0 log_2 communication volume, corresponding to about 16× compression over dense feature transmission. On DAIR-V2X, INTACT achieves 43.8 AP50 under challenging real-world conditions.

Video language models process videos as dense visual-token sequences with substantial representational redundancy. Compressing these sequences is therefore essential for reducing the visual-token burden on language-model decoding. The central challenge is to preserve visual information dispersed across frames under such compression. To this end, we introduce Aggregating Visual Information with Optimal Transport (AVIOT), which casts video token compression as transporting a dense empirical measure of frame observations onto a compact target measure. The resulting source-to-target coupling induces a distribution over source observations for each target support, directly specifying how the compressed video representation is constructed. We further adapt this construction along task and spatial axes. Question conditioning modulates the transport cost between source frames and target supports, while influencing how many supports are allocated to each temporal segment, thereby directing representation capacity toward question-relevant content. At multiple spatial granularities, AVIOT computes region-specific temporal transport plans and adaptively fuses the representations they yield, allowing different regions within the same compact representation to draw from different moments. Evaluations across varying compression ratios show that AVIOT matches or outperforms the uncompressed baseline on multiple video-understanding benchmarks while retaining strong performance at higher compression ratios.

Low-light condition significantly impair object detection performance due to severe image degradation. Existing methods fail to effectively reduce the impact of degradation, thus lacking clean detection-specific knowledge. To overcome these challenges, we propose a teacher-student model(TSM) with Clean Feature Distillation(CFD). Specifically, we design a Bridge Module(BM) to bridge the feature gap between effective image content reconstruction and detection, transferring pixel-level features to semantic-level detection features. Moreover, we propose Clean Feature Distillation (CFD) to distill degradation-free and clean features from a teacher model trained on clean images to a student model, thereby enhancing detection performance under low-light conditions. Finally, we introduce a more generalizable Low-light Image Degradation Synthesis Pipeline(LIDSP) to simulate a wide range of low-light degradation scenarios. Experimental results demonstrate that our approach outperforms existing methods, significantly boosting object detection performance under low-light conditions.

With the rapid advancement of multimodal large models and generative artificial intelligence(AI),the para-digms,understanding,and generation of image and video acquisition are undergoing profound transformations.In recent years,new-generation AI systems represented by vision-language pretraining models and diffusion-based generative models have achieved remarkable progress in semantic alignment,cross-modal understanding,and high-fidelity content genera-tion.By leveraging large-scale data and powerful representation learning capabilities,these models have significantly enhanced the performance and flexibility of visual intelligence systems,promoting their widespread adoption in intelligent security,content creation,industrial inspection,and public governance.Simultaneously,the increasing capability and deployment of visual intelligence systems have exposed a series of security risks and governance challenges,which have become increasingly prominent and cannot be disregarded.From the perspective of image and video understanding,existing visual models are frequently required to operate in complex and open-world environments that are characterized by dynamic scenes,background clutter,illumination variation,viewpoint changes,and long-tailed event distributions.In such sce-narios,the cost of obtaining large-scale,fine-grained annotations is prohibitively high,leading many practical systems to rely on limited supervision or weak labels.Although large pretrained models exhibit strong generalization capability,they still suffer from misclassification,semantic bias,and insufficient robustness when faced with domain shift,distribution mismatch,and unseen abnormal patterns.These limitations are particularly evident in safety-critical applications,where incorrect predictions or unstable behavior may result in serious consequences.Therefore,improving the reliability,robust-ness,and interpretability of image and video understanding systems has become a central topic in visual security research.In this context,anomaly detection has emerged as a core task for understanding security,because it aims to identify rare,unexpected,or abnormal events from complex visual data.Existing anomaly detection methods can be broadly categorized into fully supervised,semi-supervised,weakly supervised,and unsupervised paradigms in accordance with the availability and granularity of annotations.Fully supervised approaches rely on precise frame-level or pixel-level labels and typically achieve strong performance under controlled conditions,but their scalability and generalization capability are limited in real-world scenarios.Semi-supervised and unsupervised methods,which assume access only to normal samples during training,attempt to model normal patterns through reconstruction,prediction,or one-class learning,and detect anomalies as deviations from learned normality.Weakly supervised approaches,which are frequently formulated under the multiple instance learning framework,achieve a balance between annotation cost and detection performance,but still face chal-lenges in accurate temporal localization and the semantic interpretation of anomalies.With the emergence of vision-language large models,recent studies have begun exploring new paradigms for anomaly detection and visual understanding security.By leveraging pretrained cross-modal representations and natural language supervision,vision-language models enable zero-shot and few-shot anomaly detection,reducing reliance on task-specific annotations.Open-vocabulary anomaly recognition further allows models to detect and describe abnormal events beyond a fixed set of predefined categories,improving flexibility in open-world environments.In addition,explainable anomaly detection methods based on cross-modal alignment and attention mechanisms provide semantic-level interpretations for detected anomalies,enhancing trans-parency and trustworthiness.These advances indicate a clear trend toward a more general,scalable,and interpretable understanding of security frameworks.From the perspective of image and video generation,recent progress in generative adversarial networks(GANs)and diffusion models(DMs)has considerably improved the realism and controllability of syn-thesized visual content.GAN-based methods introduce adversarial learning mechanisms to produce visually plausible samples,while DMs further enhance generation quality and training stability through iterative denoising processes.By building upon these foundations,modern text-to-image and text-to-video generation systems integrate large vision-language models to achieve fine-grained semantic control,enabling the generation of complex scenes that closely resemble real-world data.These developments have brought significant benefits to creative industries and visual content production,but they have also amplified security risks associated with the misuse of generative technologies.High-quality synthetic images and videos can be maliciously exploited for deepfake generation,false information dissemination,identity impersonation,and privacy infringement,posing direct threats to social trust and public security.Consequently,generation security has become an essential component of image and video security research.Existing deepfake detection methods have evolved alongside generative models and can be roughly divided into several categories,including approaches based on visual arti-facts,frequency-domain characteristics,temporal consistency,and semantic coherence.While early methods focused on detecting low-level inconsistencies introduced by generation algorithms,recent approaches increasingly emphasize higher-level semantic and temporal modeling to cope with the rapid improvement of generative quality.In addition to algorithmic research,security issues associated with image and video generation have also elicited growing attention in policy regula-tion and engineering practice.Detection systems are being integrated into real-world platforms to support content modera-tion,authenticity verification,and risk assessment.Meanwhile,regulatory frameworks and technical guidelines are gradu-ally being established to govern the responsible use of generative models.Such effort highlights the necessity of combining technical solutions with governance mechanisms to address the challenges posed by generative visual technologies.Finally,despite substantial progress,image and video security in the era of large models still faces several open challenges.For understanding security,improving robustness under complex environmental changes,achieving precise temporal and spa-tial localization of anomalies,and enhancing semantic interpretability remain as key research problems.For generation security,developing generalizable deepfake detection methods that can adapt to rapidly evolving generative models remains an open issue.Moreover,balancing model capability,usability,and security constraints requires further exploration.By systematically reviewing existing research from the perspectives of understanding security and generation security,this study aims to provide a structured overview of the current landscape and offer insights into future research directions for image and video security in the era of large models.

Validating autonomous driving systems requires diverse scenarios, yet real-world data collection is biased and costly. Editing existing driving logs offers a scalable solution, but simultaneously editing objects and ego-trajectory—termed unified editing—remains challenging. Current methods face an inherent dilemma: generative flexibility for object editing and physical precision for trajectory control. To address this, we introduce SceneDirector, a diffusion-based framework that bridges explicit geometry and generative priors. For explicit geometry, we leverage LiDAR-guided depth completion to construct dense scene geometry and integrate editable 3D assets to form a Unified Geometric Scaffold, providing rigorous structural guidance for unified editing. To leverage generative priors, we encode the source video into a Static Texture Bank to provide rich appearance context. Our proposed Mask-Gated Reference Attention bridges these modalities. Guided by a geometric uncertainty metric, this mechanism dynamically regulates the interaction between the scaffold and the bank—preserving reliable geometry while adaptively injecting textures for semantic refinement. Extensive evaluations demonstrate that SceneDirector outperforms state-of-the-art methods in both controllability and visual quality.

During spacecraft launch, flight, and docking operations, monitored images often suffer from spatially variant blur caused by atmospheric turbulence, defocusing, and relative motion. To address this challenge, we propose a novel unsupervised deblurring framework tailored specifically for spacecraft imagery. Our approach incorporates three key innovations: First, we design a detail-preserving local region selection strategy based on multi-scale morphological gradients with adaptive thresholding, which optimizes regions for blur kernel estimation. Second, we define a blur kernel error term and integrate it into the degradation model, introducing explicit error correction constraints into the alternating iterative minimization process. Third, we incorporate Shearlet transform regularization to enhance the recovery of fine local details. Experimental results demonstrate that our method significantly outperforms state-of-the-art unsupervised techniques and even surpasses several advanced deep learning approaches in preserving complex structural details under spatially variant degradation. Our code and data are available at https://github.com/bsfsf/Image_deblur.

Transferring vision-language knowledge from pretrained multimodal foundation models to various downstream tasks is a promising direction. However, most current few-shot action recognition methods are still limited to a single visual modality input due to the high cost of annotating additional textual descriptions. In this paper, we develop an effective plug-and-play framework called CapFSAR to exploit the knowledge of multimodal models without manually annotating text. To be specific, we first utilize a captioning foundation model (i.e., BLIP) to extract visual features and automatically generate associated captions for input videos. Then, we apply a text encoder to the synthetic captions to obtain representative text embeddings. Finally, a visual-text aggregation module based on Transformer is further designed to incorporate cross-modal spatio-temporal complementary information for reliable few-shot matching. In this way, CapFSAR can benefit from powerful multimodal knowledge of pretrained foundation models, yielding more comprehensive classification in the low-shot regime. Extensive experiments on multiple standard few-shot benchmarks demonstrate that the proposed CapFSAR performs favorably against existing methods and achieves state-of-the-art performance. The code will be made publicly available.

Recent advancements in weakly-supervised video anomaly detection have achieved remarkable performance by applying the multiple instance learning paradigm based on multimodal foundation models such as CLIP to highlight anomalous instances and classify categories. However, their objectives may tend to detect the most salient response segments, while neglecting to mine diverse normal patterns separated from anomalies, and are prone to category confusion due to similar appearance, leading to unsatisfactory fine-grained classification results. Therefore, we propose a novel Disentangled Semantic Alignment Network (DSANet) to explicitly separate abnormal and normal features from coarse-grained and fine-grained aspects, enhancing the distinguishability. Specifically, at the coarse-grained level, we introduce a self-guided normality modeling branch that reconstructs input video features under the guidance of learned normal prototypes, encouraging the model to exploit normality cues inherent in the video, thereby improving the temporal separation of normal patterns and anomalous events. At the fine-grained level, we present a decoupled contrastive semantic alignment mechanism, which first temporally decomposes each video into event-centric and background-centric components using frame-level anomaly scores and then applies visual-language contrastive learning to enhance class-discriminative representations. Comprehensive experiments on two standard benchmarks, namely XD-Violence and UCF-Crime, demonstrate that DSANet outperforms existing state-of-the-art methods.

Text-Image Person Re-Identification (TI-ReID) aims to retrieve a target individual from an image gallery based on auxiliary textual descriptions. Current approaches face two major limitations: (1) global matching methods struggle to capture fine-grained attribute relations; and (2) textual descriptions are treated as a one-directional auxiliary modality and are often dominated by global visual representations. To address these challenges, we propose A2HA (Attribute-Aware Hierarchical Alignment), a framework designed for precise cross-modal alignment through hierarchical optimization. Specifically, we present a dynamically weighted Hierarchical Alignment (HA) framework that performs coarse-to-fine optimization across instance-, identity-, and attribute-level representations, enabling a unified modeling of global and fine-grained semantics. Rather than treating objectives as an isolated combination of existing losses, this design leverages attribute-enriched benchmarks to enable coordinated multi-level semantic alignment within a single integrated framework. To this end, we propose a dual-dimensional and dual-directional attribute-pairing strategy, capturing semantic and modal consistency while enabling bidirectional interaction between textual and visual attributes. Furthermore, we extend three widely-used benchmarks by incorporating attribute-aware segmented images and annotations, constructed with large language models and segmentation models, resulting in attribute-enriched versions of CUHK-PEDES, ICFG-PEDES, and RSTPReid. Experimental results show that A2HA achieves improvements of +5.02%, +4.68%, and +8.98% in mean Average Precision (mAP) on the three datasets, respectively. The code will be released at https://github.com/ziyuzhu452-create/A2HA.

Weakly Supervised Incremental Learning for Semantic Segmentation (WILSS) aims to segment novel classes while avoiding catastrophic forgetting based on image-level labels. However, WILSS faces two challenges. First, since features of all classes share a finite latent space, excessively dispersed base-class features limit the capacity for novel classes. Second, the limited information provided by image-level labels inevitably leads to low-quality pseudo labels, which not only hinder the learning of novel classes but also exacerbate catastrophic forgetting. To address these issues, we propose two strategies: Feature Distribution Regularization (FDR) and Pseudo Label Refinement (PLR), which are applied on the base and incremental steps, respectively. FDR employs Gaussian Mixture Models (GMMs) to model the feature distribution of each base class and enforces each feature to approach the center of its corresponding class’s GMMs, thereby making the feature distribution of base classes more compact and reserving sufficient space for novel classes. PLR improves pseudo labels from two perspectives. First, it proposes a Prototype Classification Loss (Lpcl) to ensure each prototype is correctly assigned to its corresponding class, which is beneficial for enlarging inter-class differences and helps mitigate the inter-class confusion in pseudo labels. Second, it leverages the predictions of the EMA model to adaptively refurbish the pseudo labels, as the EMA model is inherently robust to noise. The effective integration of these two methods not only facilitates the learning of novel classes but also helps mitigate catastrophic forgetting. Extensive experiments conducted on the PASCAL and COCO datasets validate the effectiveness of our method.

Translating novels into films poses a grand challenge for generative artificial intelligence, requiring conversion of abstract literary prose into long-form, multi-scene visual narratives. While current video generation models excel at short, single-scene clips within narrow temporal and spatial contexts, novel-to-film generation operates in a more complex regime, demanding long-duration content across diverse scenes with dynamically evolving entity states. To address this, we formalize novel-to-film generation as dynamic cinematic world modeling, decomposed into two phases: construction, which grounds abstract, underspecified literary narratives into concrete, stateful, and persistent world entities; and evolution, which governs how these entities dynamically update under plot progression to maintain causal consistency across scenes. We propose FilmWorld, an end-to-end agentic system where two groups of specialized agents collaborate to instantiate these phases. Construction-side agents perform narrative structured translation, world entity state modeling with visual anchoring, and state-driven shot planning, progressively projecting literary language into a cinematic blueprint. Evolution-side agents perform state-anchored visual generation, cross-shot dynamic state propagation, and closed-loop state verification to maintain causal consistency and visual coherence. To address the evaluation gap in long-form generation, we introduce FilmEval, a systematic evaluation framework that couples a difficulty-graded benchmark of 15 representative novels with an automated protocol of nine objective metrics spanning three dimensions: cinematic presentation, film consistency, and novel fidelity. Experiments demonstrate that FilmWorld consistently outperforms state-of-the-art video generation agent systems, with particularly pronounced improvements in narrative fidelity and cross-scene consistency.

Recent studies in the Text-Image Person Re-Identification (TI-ReID) tend to employ Transformers as their backbones. However, it has been studied that Transformers will bring attention noise to the assignment. To address the specific challenges encountered when applying Transformers to TI-ReID tasks, we design the network with Prioritized and Reweighted Tokens (PaRT-Net). It not only mitigates the issue of important tokens being underweighted, but also deviates from the conventional binary information processing paradigm of traditional TI-ReID methods, offering a more nuanced approach that recognizes tokens exist on a continuum of importance. Firstly, we present the Importance-Guided Token Prioritization (IGTP) module to explicitly prioritize discriminative features. Additionally, we propose the Sub-Prioritized Information Mining (SPIM) to extract complementary identity cues from tokens with sub-priority. We further put forward Hierarchical Mutual Alignment (HMA) to enhance the alignment between global and local prioritized features through mutual information theory. We conduct extensive experiments on the CUHK-PEDES, ICFG-PEDES, and RSTPReid datasets. Our method demonstrates significant improvements across all five evaluation metrics on these datasets. Specifically, our PaRT-Net achieves 83.98%, 67.71%, and 63.75% in Rank-1, and 71.25%, 42.21 %, and 52.82% in mAP, respectively. The code will be released at https://github.com/runminwang/PaRT-Net.

Multi-focus image fusion (MFIF) aims to generate an all-in-focus image from multiple images of the same scene focused at different regions. Most existing deep learning-based methods lack explicit interaction between the source images, which limits their performance and interpretability. This paper presents a novel Clarity Contrast and Similarity Selection Network (CSNet), to bridge direct information exchange for MFIF. Specifically, by contrasting the clarity differences between source images within our proposed Clarity Contrast Attention Module (CCAM), we mutually enhance sharp features while suppressing blurry ones. This allows us to identify the exactly focused regions in each source and locate the focused-defocused boundaries. Moreover, the Defocus Spread Effect (DSE) degrades pixels in all source images around the boundaries. To further refine these ambiguous areas, we introduce a Similarity Selection Strategy, which reconstructs an initial clear image from source images and selects optimal pixels by comparing the similarity among them. Through this interactive approach, CSNet effectively preserves focused regions as well as recovering natural boundaries to fuse an all-in-focus output. Extensive experiments demonstrate that our method achieves state-of-the-art performance both quantitatively and qualitatively. Our code is available on Github: https://github.com/ZYC-HUST/CSNet.

Visual Question Answering (VQA) bridges vision and language to answer image-based questions, enhancing applications in healthcare, education, and human-computer interaction. VQA models often exhibit a strong language-prior bias, disproportionately relying on spurious question-answer correlations while neglecting critical visual information. In this paper, we reinterpret the language bias and vision bias issues in VQA through the lens of both label shift and covariate shift, proposing a novel Mitigating Distribution Shift via Adaptive Reweighting (MDSAR) framework to mitigate these challenges. Specifically, we introduce two complementary strategies: (1) a label-shift adaptation mechanism that dynamically reweights question categories based on their distributional characteristics, and (2) a feature-density-aware reweighting that adjusts sample weights based on feature sparsity patterns in the dataset. By integrating these adaptive reweighting mechanisms, we formulate an enhanced loss function that effectively mitigates language bias and improves model robustness, ensuring that performance improvements on out-of-distribution (OOD) data do not come at the cost of in-distribution (ID) performance. Extensive experiments on five baseline models (i.e., UpDn, SAN, LXMERT, ViLBERT, and BLIP-2) validate the effectiveness of our approach, achieving significant performance improvements on the challenging VQA-CP v1, VQA-CP v2, VQA v2, VQA-CE, and SLAKE-CP benchmarks, and achieving state-ofthe-art performance in terms of overall metrics. Beyond VQA, the general principle of MDSAR can be readily extended to other vision-language and multimodal tasks, highlighting its broader applicability. The code will be released at https://github.com/songxdr3/MDSAR.

The presence of noise in acquired data invariably leads to performance degradation in cross-modal matching. Unfortunately, obtaining precise annotations in the multimodal field is expensive, which has prompted some methods to tackle the mismatched data pair issue in cross-modal matching contexts, termed as noisy correspondence. However, most of these existing noisy correspondence methods exhibit the following limitations: a) the problem of self-reinforcing error accumulation, and b) improper handling of noisy data pair. To tackle the two problems, we propose a generalized framework termed as Rank corrElation and noisy Pair hAlf-replacing wIth memoRy (REPAIR), which benefits from maintaining a memory bank for features of matched pairs. Specifically, we calculate the distances between the features in the memory bank and those of the target pair for each respective modality, and use the rank correlation of these two sets of distances to estimate the soft correspondence label of the target pair. Estimating soft correspondence based on memory bank features rather than using a similarity network can avoid the accumulation of errors due to incorrect network identifications. For pairs that are completely mismatched, REPAIR searches the memory bank for the most matching feature to replace one feature of one modality, instead of using the original pair directly or merely discarding the mismatched pair. We conduct experiments on three cross-modal datasets, i.e., Flickr30K, MSCOCO, and CC152K, proving the effectiveness and robustness of our REPAIR on synthetic and real-world noise.

Multi-line structured light is an essential technique for high-precision 3D measurement, where the accurate extraction of laser stripe centers directly determines reconstruction accuracy. On highly reflective metal surfaces, however, specular reflections distort the ideal intensity distribution of the laser stripes, creating bright artifacts and degrading stripe fidelity. These distortions cause deviations in center localization or even extraction failure, posing a significant challenge to precision measurement. To address these problems, this paper proposes a robust laser stripe extraction and restoration framework for reflective metal surfaces. First, a coarse-to-fine region of interest extraction strategy is developed, which reliably identifies genuine laser stripes and eliminates specular reflection artifacts by combining stripe saliency with the geometric consistency of laser projection. Then, based on the refined stripe regions, the Smooth Gaussian Boundary Tensor Voting method is adopted. By applying a compact spatial support constraint, this method effectively suppresses long-range interference while enhancing and propagating structural information along the principal stripe direction, thereby strengthening stripe continuity and restoring broken stripe segments. Experimental results confirm that the proposed framework can deliver higher accuracy and improved stability in center extraction for both straight and curved laser stripes, providing a reliable solution for high-precision 3D measurement of reflective metal surfaces.

Understanding the structure of characters is crucial for recovering clear and readable high-resolution scene text images in Scene Text Image Super-Resolution (STISR). Recently, many existing STISR methods inject the character structure information implicit in the recognition priors into the super-resolution network to guide the super-resolution process, thereby facilitating the generation of more legible text images. However, the recognition priors obtained from low resolution are inaccurate, which means that directly embedding these priors into the network easily misleads the super-resolution process. To address this problem, we draw inspiration from Masked Image Modeling (MIM) and propose the Mask Structure Inference Network (MSINet), which can generate scene text images with accurate character structures without directly embedding recognition priors. To make STISR compatible with MIM, we also propose a Mask-and-Inference Paradigm (MIP), which consists of a mask image pre-training stage for character structure learning and a fine-tuning stage for character structure inference. In addition, a novel mask strategy named Text Confidence Mask (TCM) is proposed to avoid recovery errors by masking legible character regions. With MIP and TCM, MSINet impressively improves the clarity and readability of the degraded scene text images. Specifically, MSINet-B outperforms recent state-of-the-art methods by about +3.7% on the TextZoom and average +3.6% on six manually degraded scene text recognition datasets in recognition accuracy. The code will be released at https://github.com/Yuanssr/MSINet.

Recent unified models integrate understanding experts (e.g., LLMs) with generative experts (e.g., diffusion models), achieving strong multimodal performance. However, recent advanced methods such as BAGEL and LMFusion follow the Mixture-of-Transformers (MoT) paradigm, adopting a symmetric design that mirrors one expert to another for convenient initialization and fusion, which remains suboptimal due to inherent modality discrepancies. In this work, we propose HBridge, an asymmetric H-shaped architecture that enables heterogeneous experts to optimally leverage pretrained priors from their respective Recent unified models integrate understanding experts (e.g., LLMs) with generative experts (e.g., diffusion models), achieving strong multimodal performance. However, recent advanced methods such as BAGEL and LMFusion follow the Mixture-of-Transformers (MoT) paradigm, adopting a symmetric design that mirrors one expert to another for convenient initialization and fusion, which remains suboptimal due to inherent modality discrepancies. modality domains. Unlike prior dense fusion strategies that straightforwardly connect all layers between experts via shared attention, HBridge selectively bridges intermediate layers, reducing over 40% attention sharing, which improves efficiency and enhances generation quality. Shallow and deep layers, which capture modality-specific representations, are decoupled, while mid-layer bridging promotes semantic alignment. To further strengthen cross-modal coherence, we introduce semantic reconstruction tokens that explicitly guide the generative expert to reconstruct visual semantic tokens of the target image. Extensive experiments across multiple benchmarks demonstrate the effectiveness and superior performance of HBridge, establishing a new paradigm for unified multimodal generation.

Mixture-of-Experts (MoE) architectures have emerged as a powerful paradigm for scaling diffusion models in visual generation. Recent advancements have focused on adaptively allocating computational resources across diverse tokens to improve efficiency and performance. However, we identify a routing assignment problem in existing diffusion MoE frameworks: the router fails to accurately allocate more computational resources to salient tokens. Our analysis attributes this failure to the router's reliance on noise-corrupted latent features throughout the denoising process. Such stochastic noise obscures the critical structural and textural information, thereby preventing the router from effectively distinguishing salient tokens. To address this, we propose SharpMoE, a post-training framework with a saliency-harnessing accurate routing mechanism, which utilizes clean latent features as a noise-free guidance signal for routing. By bypassing the noise-distorted inputs, SharpMoE provides the router with clear saliency guidance, enabling the identification of salient tokens even in high-noise stages. Furthermore, we introduce a trajectory routing loss to constrain the compute allocation throughout the multi-step denoising trajectory, ensuring precise resource allocation along the generation rollout. Extensive experiments demonstrate that SharpMoE serves as a versatile, plug-and-play solution that further enhances the pretrained, converged MoE models, achieving state-of-the-art performance in visual generation.

The rapid growth of modern training datasets has significantly increased computational cost, motivating dataset pruning(DP) methods which retain only a subset of informative samples to reduce training cost. Existing pruning criteria typically rely on either intrinsic signals that assess samples independently or extrinsic signals that promote diversity via pairwise relations. While effective in their own specific regimes, each captures only one aspect of sample utility and lacks robustness across different pruning ratios or data distribution. In this work, we present a unified graph-based DP framework. By modeling the dataset as a weighted graph, where node weights encode intrinsic value and edge weights encode extrinsic value, DP can be cast as a Maximum Weight Clique Problem (MWCP). Although MWCP is NP-hard, its structure admits a principled greedy solution based on sample-wise marginal gains. Under a few mild and interpretable conditions, we further prove that this unified objective enjoys a formal approximation guarantee, which applies to a broad family of importance metrics and provides practical design guidelines. Extensive experiments demonstrate that the proposed method outperforms existing pruning methods while substantially reducing training cost. On ImageNet-1k with ResNet-50, our method reduces training time by over 40\% without sacrificing accuracy.