Selected work

Representative Papers

HEDGE: Heterogeneous Ensemble for Detection of AI-GEnerated Images in the Wild

Apr 03, 2026

This work addresses the robustness challenges in detecting AI-generated images under real-world conditions, where rapid model evolution and diverse distortions degrade detector performance. To this end, the authors propose a three-way heterogeneous ensemble architecture that introduces structured heterogeneity across backbone networks (DINOv3 and MetaCLIP2), training strategies, and input resolutions. The framework incorporates a high-resolution, fine-grained forensic branch alongside progressive data augmentation and features a lightweight dual-gating mechanism for adaptive logit-space fusion. Evaluated in the NTIRE 2026 Robust AIGC Detection Challenge, the method secured fourth place and achieves state-of-the-art performance across multiple benchmarks, demonstrating significantly enhanced generalization and robustness.

0 citationsRead paper

LOGER: Local--Global Ensemble for Robust Deepfake Detection in the Wild

Apr 03, 2026

This work addresses the limited robustness of deepfake detection in real-world scenarios, where diverse forgery techniques and image degradations obscure manipulation traces. To this end, the authors propose a local–global decoupled modeling framework: a global branch leverages multi-scale heterogeneous vision foundation models to capture holistic anomalies, while a local branch employs multiple instance learning with top-k aggregation to focus on suspicious regions. Dual-level supervision is introduced to preserve discriminative power, and predictions from both branches are fused in logit space. By combining heterogeneous backbones and multi-granularity supervision, the method effectively decorrelates errors and mitigates the dilution of forgery signals by authentic image regions. The approach achieved second place in the NTIRE 2026 Robust Deepfake Detection Challenge and demonstrates strong generalization and robustness across multiple public benchmarks.

0 citationsRead paper

PosterVerse: A Full-Workflow Framework for Commercial-Grade Poster Generation with HTML-Based Scalable Typography

Jan 07, 2026arXiv.org

This work proposes the first end-to-end, commercial-grade framework for automatic Chinese poster generation, addressing limitations in design coherence, text rendering accuracy, and business adaptability that hinder existing methods from balancing high-density information with visual appeal. The framework comprises a three-stage pipeline: blueprint generation, background synthesis, and unified HTML-driven layout and text rendering. Key innovations include PosterDNA—the first HTML-formatted dataset of Chinese commercial posters—fine-tuned large language models (LLMs) for design element extraction, a tailored diffusion model for background generation, and a multimodal LLM (MLLM)-powered, extensible HTML rendering engine that effectively resolves challenges in rendering small-font, high-density text. Experiments demonstrate that the generated posters meet commercial standards in visual attractiveness, text alignment precision, and layout customizability, significantly outperforming current approaches.

0 citationsRead paper

DocIQ: A Benchmark Dataset and Feature Fusion Network for Document Image Quality Assessment

Sep 21, 2025

Document image quality assessment (DIQA) suffers from a lack of large-scale, high-quality subjective datasets and dedicated no-reference models. Method: We introduce DIQA-5000—the first large-scale subjective DIQA dataset comprising 5,000 document images—and propose a lightweight no-reference model that jointly leverages multi-level visual features and document layout priors. Our approach innovates with a layout-aware downsampling mechanism to preserve structural sensitivity at low resolutions, a multi-quality-head architecture to separately model fine-grained distributions of overall quality, sharpness, and color fidelity, and a feature fusion module to synergistically optimize low-level texture and high-level semantic representations. Contribution/Results: Extensive experiments demonstrate that our method significantly outperforms state-of-the-art general-purpose image quality assessment (IQA) models on both DIQA-5000 and OCR-oriented benchmarks, validating its effectiveness, generalizability, and practical utility for real-world document processing tasks.

0 citationsRead paper

Capturing More: Learning Multi-Domain Representations for Robust Online Handwriting Verification

Aug 02, 2025

To address the insufficient robustness in distinguishing genuine and forged signatures in online handwritten signature verification (OHV), this paper proposes SPECTRUM—a novel model that, for the first time, jointly models microscopic spectral (frequency-domain) and macroscopic temporal (time-domain) features. Methodologically, it introduces a multi-scale interaction module to extract dual-domain features, employs a self-gated fusion mechanism to dynamically weight and integrate multi-granularity time-frequency representations, and establishes a joint time-frequency distance metric framework. The core innovations lie in deep coupling of time- and frequency-domain features and a learnable cross-domain feature selection mechanism. Extensive experiments on multiple standard OHV benchmarks demonstrate that SPECTRUM significantly outperforms state-of-the-art methods, validating the substantial improvement in discriminative capability enabled by multi-domain representation learning. The source code is publicly available.

0 citationsRead paper

Recent publications

Latest Papers

HEDGE: Heterogeneous Ensemble for Detection of AI-GEnerated Images in the Wild

Apr 03, 2026

This work addresses the robustness challenges in detecting AI-generated images under real-world conditions, where rapid model evolution and diverse distortions degrade detector performance. To this end, the authors propose a three-way heterogeneous ensemble architecture that introduces structured heterogeneity across backbone networks (DINOv3 and MetaCLIP2), training strategies, and input resolutions. The framework incorporates a high-resolution, fine-grained forensic branch alongside progressive data augmentation and features a lightweight dual-gating mechanism for adaptive logit-space fusion. Evaluated in the NTIRE 2026 Robust AIGC Detection Challenge, the method secured fourth place and achieves state-of-the-art performance across multiple benchmarks, demonstrating significantly enhanced generalization and robustness.

0 citationsRead paper

LOGER: Local--Global Ensemble for Robust Deepfake Detection in the Wild

Apr 03, 2026

This work addresses the limited robustness of deepfake detection in real-world scenarios, where diverse forgery techniques and image degradations obscure manipulation traces. To this end, the authors propose a local–global decoupled modeling framework: a global branch leverages multi-scale heterogeneous vision foundation models to capture holistic anomalies, while a local branch employs multiple instance learning with top-k aggregation to focus on suspicious regions. Dual-level supervision is introduced to preserve discriminative power, and predictions from both branches are fused in logit space. By combining heterogeneous backbones and multi-granularity supervision, the method effectively decorrelates errors and mitigates the dilution of forgery signals by authentic image regions. The approach achieved second place in the NTIRE 2026 Robust Deepfake Detection Challenge and demonstrates strong generalization and robustness across multiple public benchmarks.

0 citationsRead paper

PosterVerse: A Full-Workflow Framework for Commercial-Grade Poster Generation with HTML-Based Scalable Typography

Jan 07, 2026arXiv.org

This work proposes the first end-to-end, commercial-grade framework for automatic Chinese poster generation, addressing limitations in design coherence, text rendering accuracy, and business adaptability that hinder existing methods from balancing high-density information with visual appeal. The framework comprises a three-stage pipeline: blueprint generation, background synthesis, and unified HTML-driven layout and text rendering. Key innovations include PosterDNA—the first HTML-formatted dataset of Chinese commercial posters—fine-tuned large language models (LLMs) for design element extraction, a tailored diffusion model for background generation, and a multimodal LLM (MLLM)-powered, extensible HTML rendering engine that effectively resolves challenges in rendering small-font, high-density text. Experiments demonstrate that the generated posters meet commercial standards in visual attractiveness, text alignment precision, and layout customizability, significantly outperforming current approaches.

0 citationsRead paper

DocIQ: A Benchmark Dataset and Feature Fusion Network for Document Image Quality Assessment

Sep 21, 2025

Document image quality assessment (DIQA) suffers from a lack of large-scale, high-quality subjective datasets and dedicated no-reference models. Method: We introduce DIQA-5000—the first large-scale subjective DIQA dataset comprising 5,000 document images—and propose a lightweight no-reference model that jointly leverages multi-level visual features and document layout priors. Our approach innovates with a layout-aware downsampling mechanism to preserve structural sensitivity at low resolutions, a multi-quality-head architecture to separately model fine-grained distributions of overall quality, sharpness, and color fidelity, and a feature fusion module to synergistically optimize low-level texture and high-level semantic representations. Contribution/Results: Extensive experiments demonstrate that our method significantly outperforms state-of-the-art general-purpose image quality assessment (IQA) models on both DIQA-5000 and OCR-oriented benchmarks, validating its effectiveness, generalizability, and practical utility for real-world document processing tasks.

0 citationsRead paper

Capturing More: Learning Multi-Domain Representations for Robust Online Handwriting Verification

Aug 02, 2025

To address the insufficient robustness in distinguishing genuine and forged signatures in online handwritten signature verification (OHV), this paper proposes SPECTRUM—a novel model that, for the first time, jointly models microscopic spectral (frequency-domain) and macroscopic temporal (time-domain) features. Methodologically, it introduces a multi-scale interaction module to extract dual-domain features, employs a self-gated fusion mechanism to dynamically weight and integrate multi-granularity time-frequency representations, and establishes a joint time-frequency distance metric framework. The core innovations lie in deep coupling of time- and frequency-domain features and a learnable cross-domain feature selection mechanism. Extensive experiments on multiple standard OHV benchmarks demonstrate that SPECTRUM significantly outperforms state-of-the-art methods, validating the substantial improvement in discriminative capability enabled by multi-domain representation learning. The source code is publicly available.

0 citationsRead paper