Institution profile

National Cheng Kung University

Academic institutionasia · tw
Official website
Research library107linked papers
Opportunities0open roles
Selected work

Representative Papers

MVAN: Multi-View Attention Networks for Fake News Detection on Social Media

Jun 02, 2025IEEE Access

To address the challenge of detecting fake news in real-world social scenarios—where only source tweets (short texts) and retweet user structures (without comments) are available—this paper proposes a Multi-View Attention Network (MVAN), the first framework to jointly model semantic attention over short text and structural attention over propagation graphs. MVAN employs dual-path self-attention mechanisms: one to identify salient lexical cues in the source tweet, and another to detect suspicious retweeters based on propagation topology, enabling end-to-end co-learning of semantic and diffusion patterns. The model achieves both high detection accuracy and intrinsic interpretability: it outperforms state-of-the-art methods by an average of 2.5% in accuracy on two real-world datasets, while generating traceable, semantically grounded explanations for its predictions.

62 citations4 influentialRead paper

Transformer-Driven Inverse Problem Transform for Fast Blind Hyperspectral Image Dehazing

Jan 03, 2025IEEE Transactions on Geoscience and Remote Sensing

Haze degradation severely impairs the clarity and color fidelity of hyperspectral images (HSIs), yet blind dehazing remains challenging due to the absence of haze-free reference images and ground-truth annotations. Method: This paper proposes the first end-to-end blind HSI dehazing framework. It innovatively integrates inverse problem transformation (IPT) with spectral super-resolution (SSR) to automatically identify and upsample haze-free spectral bands for initial clean HSI reconstruction. Subsequently, it introduces the first spatial–spectral Transformer, leveraging global attention to jointly model non-local spatial–spectral dependencies and refine the reconstruction. Contribution/Results: The method requires neither haze-region masks nor paired training data. Evaluated on multiple AVIRIS datasets, it significantly outperforms state-of-the-art approaches—reducing chromatic distortion, improving restoration accuracy, and demonstrating strong generalization and real-time processing potential.

4 citationsRead paper

MaTe: Images are All You Need for Material Transfer via Diffusion Transformer

Oct 19, 2025IEEE International Conference on Computer Vision

This work addresses the limitations of existing diffusion-based material transfer methods, which often rely on textual guidance or complex auxiliary networks, resulting in high computational costs and difficulties in feature alignment. To overcome these challenges, the authors propose MaTe, a lightweight diffusion framework that achieves high-quality, zero-shot, and training-free material transfer without requiring text prompts or reference networks. Built upon a diffusion Transformer architecture, MaTe performs token-level image fusion in a shared latent space through a multimodal attention mechanism, deliberately avoiding redundant components such as adapters, ControlNet, inversion sampling, or model fine-tuning. Experimental results demonstrate that MaTe surpasses state-of-the-art methods in visual quality, detail alignment accuracy, and inference efficiency, significantly simplifying deployment requirements.

2 citationsRead paper

TAG-MoE: Task-Aware Gating for Unified Generative Mixture-of-Experts

Jan 12, 2026

This work addresses the challenge of task interference in unified image generation and editing models based on dense diffusion Transformers, where shared parameters struggle to reconcile conflicting objectives such as localized editing and subject-driven generation. To mitigate this, the authors propose a task-aware Mixture-of-Experts (MoE) routing mechanism that leverages hierarchical task semantic annotations and prediction alignment regularization to guide the gating network in dispatching experts according to high-level semantic intent. This approach transforms the gate from a task-agnostic executor into a semantics-driven scheduler, enabling semantically grounded expert specialization. While preserving sparse activation, the method significantly outperforms dense baselines, achieving notable improvements in generation fidelity, editing controllability, and expert specialization.

1 citationsRead paper
Recent publications

Latest Papers

A Spectral Filtering Approach to Regret Analysis of Distributed Online Control for Linear Dynamical Systems

Aug 03, 2026

This work addresses the problem of distributed online control for networked linear time-invariant systems subject to adversarial disturbances and time-varying convex costs, where each agent has access only to its local cost information. To this end, it introduces spectral control into the distributed online learning framework for the first time: each agent constructs a spectral controller by convolving historical disturbances with the leading eigenvector of a Hankel matrix derived from local observations and neighbor communication, and collaboratively updates its parameters via distributed online gradient descent. The approach establishes a regret analysis framework based on spectral parameterization and, under standard assumptions, proves a sublinear regret bound of $O(\frac{\sqrt{T}\,\text{poly}(\log T)}{\gamma^3})$, explicitly characterizing the dependence on the time horizon $T$, the system’s stability margin $\gamma$, and the network size and connectivity.

0 citationsRead paper

Reconstructing Persistent Worlds from Narratives for Narrative-Grounded Interactive Experiences

Aug 03, 2026

This work addresses the limitation of existing narrative-driven interactive experience generation methods, which often fail to model story-world consistency, leading to a disconnect between generated content and narrative context. To bridge this gap, we propose explicitly reconstructing a persistent narrative world as the central computational task, formalizing it as a structured representation encompassing entities, locations, semantic relationships, and state evolution. This unified world model jointly supports narrative understanding and interactive content generation. Integrating structured world modeling, narrative semantic parsing, and contextual reasoning, we present the first prototype system that treats persistent world reconstruction as its core objective, enabling automatic transformation from narrative text into playable tile-based environments. Evaluation through three case studies demonstrates that our approach generates semantically coherent and gameplay-consistent interactive worlds, showing promising potential for applications in AI-assisted game design and educational simulations.

0 citationsRead paper

FDIR: Harmonizing Fidelity and Human-Machine Preference in Lossy Compression Image Restoration

Jul 31, 2026

Existing lossy compressed image restoration methods struggle to simultaneously achieve high pixel-level fidelity, perceptual quality for human observers, and performance on downstream machine vision tasks, often resulting in trade-offs among these objectives. This work proposes FDIR, a two-stage framework that first recovers global semantic structure in latent space via quality-guided one-step flow matching (QO-Flow), followed by deterministic high-frequency texture reconstruction in pixel space using a flow-conditioned detail refinement module (FCDR) to effectively suppress hallucinations. FDIR is the first unified approach to jointly optimize all three goals through a decoupled design that balances their inherent conflicts, thereby avoiding excessive smoothing and distortion. Experiments demonstrate that FDIR achieves state-of-the-art or competitive performance across fidelity, perceptual quality, and machine task metrics.

0 citationsRead paper