Addressing complex meteorological processes at a fine spatial resolution requires substantial computational resources. To accelerate meteorological simulations, researchers have utilized neural networks to downscale meteorological variables derived from low-resolution simulations. Despite notable advancements, the contemporary cutting-edge downscaling algorithms are tailored to specific variables. Handling meteorological variables in isolation involves overlooking their interconnectedness, leading to an incomplete understanding of atmospheric dynamics. Additionally, the laborious processes of data collection and processing and the computational resources required to downscale individual variables are significant hurdles. Given the limited versatility of the existing models across different meteorological variables and their failure to account for intervariable relationships, this paper proposes a unified downscaling approach leveraging meta-learning, with its architecture based on the Enhanced Deep Super-Resolution (EDSR) model. The proposed framework supports downscaling across diverse variables and xsclimate datasets. It exhibits strong extensibility by generalizing to 18 unseen during training variables from CFS, S2S (CMA), and CMIP6 (CMCC-ESM2)—such as mean sea-level pressure, temperature at 850 hPa, and the 500 hPa U- and V-wind components—marking a step toward a universal downscaling solution. Experimental evidence demonstrates that the proposed model outperforms the best existing methods, such as EDSR and ClimateSD, in both quantitative and qualitative assessments.

Abstract. High-resolution reconstruction of ocean dynamics is challenging because spectral bias and the neglect of cross-variable couplings in existing super-resolution (SR) methods often lead to over-smoothed, physically inconsistent outputs, limiting their utility for eddy parameterizations. To overcome these limitations, we present a Multiphysics Super-Resolution version 1.0 (MSR v1.0) framework that jointly reconstructs eight closure-relevant diagnostics—vorticity, deformation measures, stress tensor components, and subgrid momentum forcing terms—directly from low-resolution (LR) velocity fields, consistency is maintained under a filtering scale that aligns with the definition of subgrid tendencies. Our approach integrates three key components: (1) a Dynamic Enhancement Feature (DEF) module to prioritize dynamically active regions; (2) a High-Frequency Enhancement (HFE) module that fuses spatial and spectral operators via learned gating to restore suppressed fine-scale structures such as fronts and eddy rims; and (3) a Physical Consistency Loss that aligns derivative-level structures and algebraic couplings across diagnostics. Experiments on an eddy-resolving simulation dataset across multiple ocean basins and downscaling factors show that MSR consistently outperforms strong SR baselines, yielding sharper reconstructions with improved high-wavenumber spectra and cross-variable consistency. The MSR-reconstructed diagnostics are closure-ready for low-resolution ocean models and can inform or constrain eddy parameterizations, providing a practical, spectrally selective, scale-aware surrogate for high-fidelity multiphysics diagnostics.

Positron emission tomography (PET) is an advanced nuclear imaging technique and has been widely applied in clinic. However, radiation risks associated with standard-dose PET imaging raise health concerns, whereas the quality of low-dose PET images fails to meet clinical requirements. To reduce the tracer dose while maintaining image quality, it is of great interest to estimate high-quality PET images from low-dose images. However, existing low-dose PET image denoising methods primarily focus on image data, overlooking crucial information in non-image textual data such as patients' clinical tabular and textual descriptions of general image quality. This neglect can lead to subpar denoising quality with inaccurate contexts and poor details. To address these problems, in this paper, we propose Multi-Granularity Textual Prompts, namely MGTP, to denoise low-dose PET images via an adversarial diffusion model. Different from prior methods that rely solely on image conditioning, our MGTP innovatively introduces textual prompts spanning diverse granularities to capture both high-level semantic-related contexts and low-level degradation-related details. To harmonize multi-granularity textual prompts with low-dose PET images, we design a Cross-Modality Selective Conditioning (CMSC) module, which prioritizes semantic- and detail-relevant information while eliminating irrelevant components. The resulting features are fed into diffusion model as conditions, enforcing a more controlled diffusion process. In addition, we develop a Masked Prompt Reconstruction Network (MPR-Net) to enhance the preservation of semantics and details in denoised images, mitigating distortions brought by the random noise in the diffusion process. Experiments on clinical PET data show that our method achieves the state-of-the-art performance.

Accurate and layout-consistent reconstruction remains a key challenge in indoor simultaneous localization and mapping (SLAM) due to the prevalence of planar and axis-aligned structures. Traditional visual and RGB-D SLAM methods often suffer from incomplete geometry and weak structural reasoning, while NeRF-based SLAM improves fidelity but is computationally expensive and unsuitable for real-time use. 3D Gaussian splatting offers improved efficiency but lacks structural priors, often resulting in distortions in structured scenes. To address these issues, we propose a structure-aware SLAM framework based on 2D Gaussian splatting, which provides efficient, view-consistent mapping. We introduce a lightweight regularization scheme under the Manhattan-world assumption to align Gaussian orientations and positions with dominant axes, improving layout consistency and geometric fidelity. Extensive experiments on Replica and TUM-RGBD datasets demonstrate that our method consistently outperforms existing SLAM baselines in terms of geometric accuracy and edge preservation across multiple indoor scenes.

Current Convolutional Neural Networks (CNNs) for Weakly Supervised Semantic Segmentation (WSSS) often have difficulties in discovering distinctive feature locations for each category. Therefore, the pseudo-labels generated from the expanded seed regions are typically incomplete and contain a significant amount of noise. Without additional annotations, the numerous erroneous information will potentially propagate in the segmentation network's training stage. In this work, we propose a Cross-Modal Dual Graph Reasoning (CDGR) framework to leverage both visual and language knowledge effectively. This framework can capture dependencies between the spatial and the semantic spaces, facilitating the discovery of discriminative feature locations. Specifically, we perform cross-modal graph reasoning between the visual and the language modal graphs to enhance global contextual relationships between pixels in the visual feature map. Additionally, we introduce a graph interaction attention network to thoroughly explore implicit relationships between visual and language graphs. We apply the CDGR network to generate more complete pseudo-labels for the classification network and utilize it in the segmentation network to unleash its self-correcting capabilities. Extensive experiments on the PASCAL VOC 2012 and MS COCO 2014 datasets demonstrate the effectiveness of CDGR compared to other state-of-the-art peers. Our code is provided at https://github.com/JIA-ZHANG666/CDGR

Restoring high-quality images from blurred videos is a highly challenging task, especially in severely blurred scenes. In recent years, event-based methods have achieved significant progress in video deblurring. However, the modal differences between the event and image increase the difficulty of feature fusion. Additionally, the sparsity of event makes it difficult to restore some local details. To address these issues, we propose a new video deblurring method. Firstly, we design a cross-modal collaborative attention mechanism to effectively fuse features from blurred frames and event frames, thereby deeply extracting motion information from event frames. Secondly, we utilize a diffusion model to generate spatial guiding prior feature, enhancing local details and textures. Furthermore, we propose an event-guided dynamic feature fusion module that adaptively integrates spatio-temporal information from neighboring frames. Experimental results on both synthetic and real datasets demonstrate that our method outperforms the current state-of-the-art approaches. The code is available at: https://github.com/Frank-Zhou-01/EDVDmain.

The scarcity of paired training samples and significant cross-modality disparities poses substantial challenges for visible-infrared person re-identification (VI-ReID). Existing methods primarily focus on generating middle-modality images to alleviate the data scarcity and bridge the modality gap. However, despite progress, these approaches face limitations: the auxiliary role of middle modalities remains constrained, network structures tend to be simplistic, and the design of effective loss functions is challenging. These limitations hinder models from fully leveraging the information within the middle modality. To address these issues, this paper introduces an Interactive Feature Attention Learning Network (IFALNet). The framework first employs a Middle Modality Generator (MMG) to synthesize middle-modality images, which are then processed by the feature extraction network. Within this network, a Modal Interactive Attention (MIA) module captures discriminative information regarding modality-shared features, while a Multi-scale Feature Aggregation (MFA) block focuses on modality-specific discriminative information, collectively enhancing feature representation. Furthermore, a Triple Centers Aggregation (TCA) loss guides the feature learning across visible, infrared, and middle modalities. This loss function effectively reduces the distance between modality centers, thereby eliminating modality gaps. Extensive experiments conducted on the SYSU-MM01, RegDB, and LLCM datasets demonstrate the superiority of the proposed method. Notably, compared to the recent MMN approach of the same type, IFALNet achieves significant improvements of 6.6 % in Rank-1 accuracy and 6.2 % in mAP on the challenging SYSU-MM01 dataset. Code is available at https://github.com/ZHY-tech11/IFALNet.

Multimodal large language models (MLLMs) have progressed substantially in recent years and now support strong cross modal capabilities such as visual grounding, document understanding, and multi-step reasoning. Compared with unimodal LLMs, however, MLLMs are structurally heterogeneous: the vision stack is typically compute-heavy and resolution-sensitive, while the language stack is deeper and often memory-bound. This mismatch makes distributed recipes designed for homogeneous Transformer stacks (e.g., symmetric Megatron-style partitioning) inefficient in MLLM training, leading to stage imbalance, pipeline bubbles, and expensive cross-modal boundary communication. Existing MLLM deployments usually adopt either fixed co-location or strict module separation. Fixed co-location can reduce communication for one preset split but is brittle across model scales and resolutions, whereas strict separation preserves modularity but amplifies vision–language boundary communication overhead. To address these limitations, we propose SelCo (Selective Colocation), an adaptive distributed training architecture for heterogeneous multimodal graphs. SelCo combines asymmetric partitioning, which assigns different pipeline spans to the vision-heavy and language-heavy portions of the model, with Selective Co-location, which places the high-traffic vision–language boundary within the same GPU group when doing so reduces critical-path communication. SelCo is driven by PPE (Profiler–Planner–Executor): the profiler collects in-situ runtime costs, the planner solves a mixed-integer linear program (MILP) to select hardware-aware asymmetric partitions, and the executor enforces the plan with selective co-location and schedule-safe runtime mechanisms. This design enables zero-copy transfer on the critical path while preserving correctness under asymmetric 1F1B execution. On a commodity 8× RTX 3090 cluster with CLIP ViT-L/14-336 and Qwen2.5-7B, SelCo achieves up to 4.94× higher throughput than a homogeneous baseline, demonstrating improved utilization under resource-constrained training.

Precipitation nowcasting is vital for protecting lives and economic activities, yet accurate forecasts based solely on past precipitation remain elusive. Conventional numerical weather prediction models offer a solution but incur substantial computational costs. Moreover, due to the rapid pace of climate change, long-term time series data are often inadequate for accurately addressing precipitation forecasting for extreme weather events in a short period of time, as past meteorological time series data may not accurately reflect current atmospheric conditions. There is an urgent need to rely on short-term time series for prediction tasks. Existing studies have employed Spatio-Temporal Information Transformation (STI) equations with iterative solutions for short-term time series prediction. However, the solution process involves relatively simple nonlinear operations, which are prone to cumulative errors and can result in inaccurate forecasts. In response, the present work proposes a dual encoder-decoder training framework based on the STI equation and the idea of dual learning, which can map multidimensional spatial features to the temporal prediction of future precipitation variables. This architecture addresses the limitations of inaccurate predictions for short-term time series data. Additionally, an adaptive weighted gradient loss (ADGLoss) is proposed to mitigate the prediction ambiguity caused by the extension of prediction time and rectify systematic underestimation of high-intensity precipitation regions. Leveraging the SEVIR dataset, the proposed model integrates multiple meteorological variables to generate 1 h precipitation forecasts via Vertically Integrated Liquid (VIL) estimation. Experimental results demonstrate that the STI-driven framework achieves superior predictive accuracy and reduced error rates in hourly multi-step forecasting compared to state-of-the-art deep learning benchmarks. The model effectively captures the spatio-temporal dependencies between heterogeneous meteorological variables and VIL precipitation patterns, offering a novel pathway for advancing spatio-temporal prediction tasks in climate informatics.

While single-task medical image restoration (MedIR) has achieved remarkable progress, developing a unified model that generalizes across diverse MedIR tasks remains largely unexplored, primarily due to the complex degradation patterns present in medical images across different modalities. To tackle this challenge, in this paper, we propose UniMedIR, an all-in-one framework that unifies multiple MedIR tasks within a single model through text-guided prompting and dual-domain modeling. Specifically, UniMedIR leverages large language model (LLM)enhanced descriptions to provide task-relevant guidance by explicitly encoding restoration objective, imaging modality, and degradation characteristics. To further improve prompt quality, we introduce a mixture of task experts (MoTE), which captures both task-shared anatomical structures and task-specific degradation patterns to generate adaptive and discriminative task prompts. Through this design, our UniMedIR learns common structural representations while also modeling modality-dependent degradations, enabling robust generalization across diverse tasks and modalities. Guided by these prompts, we then develop a dual-domain transformer (DuDoTrans) that performs joint spatial and frequency modeling for enhanced restoration. Within the DuDoTrans, a prompt-guided spatial attention (ProSA) recovers task-adaptive structure information in the spatial domain, while an adaptive frequency modulator (AdaFM) adaptively decomposes and refines high-frequency and low-frequency components to restore structure and texture details. In addition, we build a comprehensive benchmark across six representative MedIR tasks spanning computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET) modalities. Extensive experiments demonstrate that UniMedIR achieves the state-of-the-art performance in both quantitative and qualitative evaluations among all-in-one MedIR approaches and exhibits superior robustness to unseen settings. Code is available at https://github.com/gluucose/UniMedIR.

Face swapping aims to transplant source face identity onto a target image while preserving pose, expression, illumination, and background. Despite advances in diffusion models, existing methods suffer from identity degradation and background inconsistency. This work presents a conditional denoising diffusion framework for high-fidelity face swapping. We adopt disentangled representation learning to separate source identity and target attributes via cross-attention. A ground-truth noise-based supervision strategy enhances identity preservation, and an iterative face-background fusion mechanism improves visual coherence. Experiments on standard benchmarks show that the proposed method outperforms representative baselines in identity fidelity and visual realism.

Sleep staging is a crucial link between brain function monitoring and regulation. Current deep learning-based methods have unlocked a new paradigm for feature learning in automatic sleep staging modeling. However, most deep learning models still lack robustness to the domain-invariant features of sleep signals. Making them unable to effectively address individual differences and data imbalance in the automatic sleep staging task. In this paper, we propose a delayed adversarial optimization neural network approach that can improve the learning of domain-invariant features in the automatic sleep staging task. Specifically, we introduce the simplex equiangular tight frame (ETF) of neural collapse to depict the domain-invariant features of sleep signals. Constraining the imbalanced staging features to align with the ETF of balanced data learning. Then, to simulate the data property of individual differences among different subjects, we generate adversarial minority samples by perturbing the loss of the staging model. In addition, an ETF joint optimization strategy is established to fine-tune the staging model, achieving the domain-invariant feature gain. We validated the proposed method on MASS and Sleep_EDF datasets under a cross-subject test experiment. The test results showed that our method has significant advantages in addressing the problem of individual differences.

Accurate precipitation nowcasting plays an important role in disaster prevention and hydrometeorological applications, yet it remains highly challenging due to the complex spatiotemporal variability and multi-scale structural characteristics of precipitation systems. Existing deep learning methods are largely data-driven and often struggle to effectively exploit multi-source observations or learn physically meaningful representations. To address these limitations, this study proposes a Multi-Scale Frequency-Temporal Network (MS-FTNet) for precipitation nowcasting. The framework leverages Fourier transform-based frequency-domain modeling to achieve an interpretable multi-scale decomposition of precipitation dynamics. Specifically, low-frequency components capture large-scale stratiform patterns and their temporal evolution, while high-frequency components represent localized convective structures and abrupt variations. Building on this, a Global Feature Collaboration (GFC) module integrates global frequency-domain representations with multi-scale convolutional features, and an Adaptive Temporal Fusion (ATF) module enhances temporal dependency modeling. Experiments on the SEVIR dataset demonstrate that MS-FTNet consistently outperforms representative baseline models in terms of MSE, CSI, and LPIPS, particularly for heavy precipitation events and longer forecast lead times.

Facial Expression Recognition (FER) is crucial to human-computer interaction. Existing cross-domain FER (CD-FER) methods mainly focus on single-source closed-set scenarios, transferring knowledge from a single source domain to a target domain with identical class sets. However, CD-FER faces two real-world challenges: 1) the need to leverage information from multiple sources, leading to multi-domain shift, and 2) the necessity to recognize unseen target classes, resulting in class shift. These issues give rise to a novel and challenging task, which we define as Multi-domain Open-set FER (MO-FER). In this paper, we propose PromptEmo, a novel CLIP-based framework that leverages bilateral textual prompts to address both shifts in the MO-FER task. Leveraging the generalizability of LLM, PromptEmo constructs trainable positive prompts with LLM-generated emotion descriptions for seen classes, as well as template-derived negative prompts to enhance the reasoning for unseen classes. Then, we introduce a modal-task optimization paradigm organized from two perspectives: textual semantics and visual domains, yielding Intra-modal Space-specific Optimization (ISO) and Cross-modal Emotion-aware Interaction (CEI) strategies. ISO refines the CLIP-based textual space to ensure semantic separation between bilateral prompts and improves the latent visual space by promoting inter-domain alignment. Founded on ISO, CEI facilitates effective vision-language interactions, resulting in four joint loss terms that improve emotion recognition by shaping a domain-invariant, discriminative feature space. PromptEmo surpasses the current SOTA method by 7.7% AUC on unseen classes across four FER datasets, serving as a strong baseline for the MO-FER task.

Echocardiography plays an important role in the screening and diagnosis of cardiovascular diseases. However, automated intelligent analysis of echocardiographic data remains challenging due to complex cardiac dynamics and strong view heterogeneity. In recent years, visual language models (VLM) have opened a new avenue for building ultrasound understanding systems for clinical decision support. Nevertheless, most existing methods formulate this task as a direct mapping from video and question to answer, making them vulnerable to template shortcuts and spurious explanations. To address these issues, we propose EchoTrust, an evidence-driven Actor-Verifier framework for trustworthy reasoning in echocardiography VLM-based agents. EchoTrust produces a structured intermediate representation that is subsequently analyzed by distinct roles, enabling more reliable and interpretable decision-making for high-stakes clinical applications.

Existing deep learning-based superpixel methods typically rely on a fixed grid size to establish pixel-superpixel associations, which limits their ability to handle objects at varying scales. To address this limitation, we propose a Hierarchical Superpixel Segmentation Network (HSSN) inspired by the human visual attention mechanism. First, we introduce a Local Scale Estimation Module (LSEM) that utilizes geodesic distance to quantify pixel similarity. This module categorizes image regions into large, medium, and small scales, enabling adaptive grid initialization for objects of different sizes. Building on this, we incorporate parallel multiscale branches with auxiliary losses to enhance the representation of structural details. Furthermore, we design a Scale-adaptive Fusion Module (SFM) that integrates contextual features using spatially-aware weighting. Experimental results demonstrate that HSSN improves segmentation accuracy while maintaining computational efficiency. It also provides sharper boundary cues for downstream tasks, such as salient object detection.

Recently, the object detection methods based on few shot learning have significantly advanced the field of object detection, because the methods are easy to achieve excellent detection results when there are few training samples. However, detecting novel classes with existing methods often requires time-consuming model retraining. Moreover, the detection performance of new models for base classes may decrease. In order to solve the above problems, this paper proposes a new few-shot object detection model, which can efficiently detect the novel classes introduced without fine-tuning the model. This model enables simultaneous detection of both novel and base classes, effectively mitigating catastrophic forgetting. The model incorporates a base class detector, augmented with an additional contrastive branch to extract class representation information. This decouples object localization and classification, leading to a marked improvement in the model’s generalization performance for novel classes. In addition, we investigate the effectiveness of self-supervised and supervised contrastive losses for class-specific encodings in our framework.

Kilometre-scale daily precipitation forecasts at Sub-seasonal to Seasonal (S2S) lead times are critical for climate-sensitive sectors, but General Circulation Model (GCM) outputs are coarse and biased. Their ensemble nature also undermines traditional downscaling methods such as Quantile Mapping, which assume predefined relationships. Generative Adversarial Networks (GANs) can generate realistic precipitation fields but are unstable to train, while diffusion models struggle with mismatched coarse-to-fine training pairs and frequent zero-precipitation values. This study introduces the Conditional Diffusion Downscaling Model (CDDM), which frames S2S downscaling as conditional image translation from coarse GCM forecasts to fine-resolution observed precipitation fields. Drawing on the Brownian bridge process, CDDM modifies standard forward diffusion by terminating at a hybrid state that blends the coarse forecast with controlled Gaussian noise, rather than pure noise. The reverse process, conditioned on the raw GCM forecast, then learns a stochastic mapping to fine-scale precipitation fields. A composite loss function combining Mean Absolute Error (MAE) and relative Bias (rBias) further improves forecast accuracy.CDDM is evaluated by downscaling nine-member ACCESS-S2 ensemble forecasts from 60 km to 5 km resolution over eastern Australia. Averaged over three evaluation years and 42-day forecast horizons, CDDM consistently outperforms QM, a climatology benchmark, and a state-of-the-art GAN across probabilistic and deterministic metrics, including the continuous ranked probability score (CRPS), MAE, and rBias. It achieves improvements of at least 8.29\% in CRPS and 9.01\% in MAE, and remains robust across La Niña, neutral, and El Niño conditions.