Zero-Shot Composed Image Retrieval (ZS-CIR) aims to retrieve images by combining a reference image with a textual description, without requiring supervision from target images. While existing self-supervised methods fuse visual and textual representations, they often struggle to capture fine-grained visual details and lack robust cross-modal interactions due to inadequate alignment supervision. To address this, we propose a novel framework, Dual-granularity Image-Text Alignment (DITA). DITA employs an Adaptive Token Pooling (ATP) module to compress visual tokens into an informative set, preserving rich image details for effective text querying while eliminating redundancy. It then introduces two core contrastive objectives: (i) an image-only prompt, formed by combining template text with the pooled tokens, aligns with the reference image for its cross-modal self-awareness, and (ii) a masked composite prompt, where a caption with masked keywords is appended to the image-only prompt, interacts with synthetic hard descriptions for image-text alignment and promotes their fine-grained fusion. To rigorously evaluate DITA, we conduct extensive experiments on FashionIQ, CIRR, and CIRCO benchmarks using diverse backbones (e.g., BLIP, BLIP-2, SigLIP). The experimental results demonstrate that DITA consistently outperforms competitive self-supervised baselines. Furthermore, DITA exhibits exceptional robustness against adversarial queries and varying text complexities. With both BLIP-based backbones, DITA processes each query in about 14 ms on a single RTX 4090 GPU. Compared with OSrCIR, a representative LLM-based training-free method relying on API calls, DITA achieves about 40x lower online latency while maintaining comparable retrieval performance. These results confirm that DITA provides a scalable, robust, and highly efficient solution for real-world ZS-CIR applications.
A two-step microarc oxidation (MAO) fabricates in situ Ce-doped coatings with self-sealing and self-healing capabilities on the magnesium (Mg) alloy. The CePO4/MgO composite melt effectively seals discharge pores, while its nanocrystalline-amorphous structure enhances coating densification and corrosion resistance. Electrochemical and salt solution demonstrate that Ce-doped coatings exhibit excellent corrosion resistance. During corrosion, released Ce3+ promotes formation of Ce(OH)3/Mg(OH)2 products. Density functional theory (DFT) calculations reveal these products exhibit increased work function and weakened Cl− adsorption, synergistically forming a robust electronic barrier that suppresses corrosive attack and enables persistent self-healing protection.
The extracellular matrix (ECM) undergoes extensive architectural remodeling during tumor progression, generating nanoscale alignment cues that profoundly influence behaviors of immune cells. However, whether ECM nanoarchitecture directly programs their immune secretory functions remains poorly defined. Here, we develop a magnetically switchable hydrogel that enables reversible, on-demand modulation of nanoscale ligand alignment without altering bulk stiffness or composition. Using this platform, we uncover a mechanotransduction pathway in tumor-associated macrophages (TAMs) whereby matrix nanoalignment enhances integrin β1-dependent mechanosensing and transient receptor potential cation channel subfamily V member 2 (TRPV2)-dependent Ca2+ signaling, thereby promoting phosphoinositide 3-kinase (PI3K)–Ras-related protein 11a (Rab11a)-associated small extracellular vesicle (sEV) release by approximately tenfold. Beyond increasing vesicle output, PI3K–mechanistic target of rapamycin complex 2-associated signaling altered the particle-normalized cargo profile of sEVs, with preferential enrichment of immunoregulatory molecules, including M2-promoting cytokines and miR-181a-5p. These sEVs reprogram macrophages toward an immunosuppressive M2-like phenotype, suppress cytotoxic CD8+ T-cell infiltration, and accelerate melanoma progression in vivo. Importantly, selective disruption of the TRPV2–Rab11a axis or pharmacological attenuation of ECM remodeling restores antitumor immunity and restricts tumor growth. Together, our findings demonstrate that ECM nanoarchitecture is sufficient to regulate macrophage sEV programming in a defined matrix environment and identify TRPV2-associated mechanotransduction as a potential therapeutic vulnerability in matrix-remodeled tumors.
Litchi is popular among consumers due to its delicious taste. However, the extremely short shelf life limits its commercial value. Conventional single-gap dielectric barrier discharge (SG-DBD) treatment may provide non-thermal microbial control, but the curved geometry of spherical fruit can distort the electric-field distribution and lead to non-uniform plasma exposure. In this study, an innovative double-gap dielectric barrier discharge (DG-DBD) formed with a 3D-printed frame was introduced to improve the electric field uniformity over spherical loads. Its effectiveness was supported by electrostatic modeling electrical characterization, and preservation experiment on the spherical fresh produce like litchi. The DG configuration outperformed the single-gap (SG) configuration for decay, browning, and control of yeasts and molds. And fruit spacings of 1-2 cm produced more favorable preservation outcomes than close-packed placement. These results indicate that geometry-aware DBD treatment can improve plasma exposure characteristics and may enhance the postharvest preservation of spherical fresh produce.
Capturing the rich, hierarchical information embedded in complex multivariate time series (MTS) necessitates modeling at multiple temporal scales. However, prevalent multi-scale forecasting methods suffer from critical limitations: a uniform scaling approach that overlooks individual variable characteristics, inconsistent modeling of temporal patterns and inter-variable dependency leading to modeling spurious correlations, and a neglect of global context. To overcome these challenges, we introduce LMSNet, a Lightweight Multi-Scale graph learning framework. Specifically, LMSNet introduces a novel variable-specific scaling mechanism, which leverages frequency-domain analysis to assign a tailored analytical window to each variable based on its intrinsic dynamics, establishing a robust foundation for accurately modeling multi-scale inter-variable dependencies. Furthermore, LMSNet synergizes global and local information and performs concurrent temporal and inter-variable dependency modeling at every network layer, aided by a momentum update mechanism that refines deep-layer graph node representations to mitigate the risk of modeling spurious correlations. LMSNet establishes a new state-of-the-art on eight benchmark datasets. Crucially, it achieves a 2.94% reduction in MAE with reducing GPU memory usage by 83% and accelerating training by 60% compared to strong baseline-TimeMixer. Our source code is provided at https://github.com/05Pikachu24/TSF-LMSNet.
