Selected work

Representative Papers

HeadsUp! High-Fidelity Portrait Image Super-Resolution

Oct 10, 2025

Existing portrait super-resolution methods typically adopt a hybrid “face-specific + generic model” strategy; however, inconsistent training objectives between the two components often introduce severe artifacts at facial-background boundaries, degrading visual realism. To address this, we propose HeadsUp, an end-to-end, single-step diffusion framework that unifies holistic portrait reconstruction. First, we incorporate explicit facial-region supervision to enhance local detail fidelity. Second, we design a reference-guided identity consistency restoration mechanism to preserve subject identity. Third, we construct PortraitSR-4K—a high-quality, 4K-resolution portrait dataset—to support both training and rigorous evaluation. Extensive experiments demonstrate that HeadsUp achieves state-of-the-art performance across multiple benchmarks, significantly suppressing boundary artifacts while maintaining strong generalizability to both generic image super-resolution and aligned face super-resolution tasks.

0 citationsRead paper

TrueSkin: Towards Fair and Accurate Skin Tone Recognition and Generation

Sep 13, 2025

Skin tone recognition and generation face core challenges including data scarcity, insufficient model robustness, and algorithmic bias—particularly systematic misclassification of intermediate skin tones and interference from irrelevant attributes (e.g., hairstyle, background). To address these, we introduce TrueSkin, the first systematically annotated, multi-condition skin tone benchmark comprising 7,299 real-world images across six tone categories, captured under diverse lighting conditions, viewpoints, and imaging devices. Leveraging TrueSkin, we develop a supervised classification framework and a fine-tuning pipeline for generative models. Experiments demonstrate that our recognition model achieves over 20% higher accuracy than state-of-the-art methods; our generation framework significantly improves target skin tone fidelity while mitigating bias induced by confounding attributes. Critically, TrueSkin reveals, for the first time, systematic bias of large foundation models toward intermediate skin tones. It establishes a reproducible benchmark and methodological foundation for fair, robust skin tone perception and synthesis.

0 citationsRead paper

4KAgent: Agentic Any Image to 4K Super-Resolution

Jul 09, 2025

This paper introduces the first autonomous agent paradigm tailored for low-level vision tasks, addressing the long-standing challenge of generalizable super-resolution (SR) reconstruction from arbitrarily severely degraded low-resolution images (e.g., 256×256) to 4K+ outputs. Methodologically, it proposes a perception–restoration dual-agent architecture: a perception agent leverages vision-language models to interpret image content and invoke domain-specific quality evaluators; a restoration agent employs recursive execute–reflect loops and a quality-driven mixture-of-experts strategy to dynamically select and orchestrate specialized submodules—including SR, denoising, and facial restoration—with a dedicated facial restoration pipeline ensuring fidelity in critical regions. Evaluated across 11 task categories and 26 benchmarks, the method achieves state-of-the-art performance on diverse modalities (natural, medical, and satellite imagery), significantly improving perceptual quality (LPIPS ↓) and fidelity (NIQE ↓, FID ↓).

0 citationsRead paper

Recent publications

Latest Papers

HeadsUp! High-Fidelity Portrait Image Super-Resolution

Oct 10, 2025

Existing portrait super-resolution methods typically adopt a hybrid “face-specific + generic model” strategy; however, inconsistent training objectives between the two components often introduce severe artifacts at facial-background boundaries, degrading visual realism. To address this, we propose HeadsUp, an end-to-end, single-step diffusion framework that unifies holistic portrait reconstruction. First, we incorporate explicit facial-region supervision to enhance local detail fidelity. Second, we design a reference-guided identity consistency restoration mechanism to preserve subject identity. Third, we construct PortraitSR-4K—a high-quality, 4K-resolution portrait dataset—to support both training and rigorous evaluation. Extensive experiments demonstrate that HeadsUp achieves state-of-the-art performance across multiple benchmarks, significantly suppressing boundary artifacts while maintaining strong generalizability to both generic image super-resolution and aligned face super-resolution tasks.

0 citationsRead paper

TrueSkin: Towards Fair and Accurate Skin Tone Recognition and Generation

Sep 13, 2025

Skin tone recognition and generation face core challenges including data scarcity, insufficient model robustness, and algorithmic bias—particularly systematic misclassification of intermediate skin tones and interference from irrelevant attributes (e.g., hairstyle, background). To address these, we introduce TrueSkin, the first systematically annotated, multi-condition skin tone benchmark comprising 7,299 real-world images across six tone categories, captured under diverse lighting conditions, viewpoints, and imaging devices. Leveraging TrueSkin, we develop a supervised classification framework and a fine-tuning pipeline for generative models. Experiments demonstrate that our recognition model achieves over 20% higher accuracy than state-of-the-art methods; our generation framework significantly improves target skin tone fidelity while mitigating bias induced by confounding attributes. Critically, TrueSkin reveals, for the first time, systematic bias of large foundation models toward intermediate skin tones. It establishes a reproducible benchmark and methodological foundation for fair, robust skin tone perception and synthesis.

0 citationsRead paper

4KAgent: Agentic Any Image to 4K Super-Resolution

Jul 09, 2025

This paper introduces the first autonomous agent paradigm tailored for low-level vision tasks, addressing the long-standing challenge of generalizable super-resolution (SR) reconstruction from arbitrarily severely degraded low-resolution images (e.g., 256×256) to 4K+ outputs. Methodologically, it proposes a perception–restoration dual-agent architecture: a perception agent leverages vision-language models to interpret image content and invoke domain-specific quality evaluators; a restoration agent employs recursive execute–reflect loops and a quality-driven mixture-of-experts strategy to dynamically select and orchestrate specialized submodules—including SR, denoising, and facial restoration—with a dedicated facial restoration pipeline ensuring fidelity in critical regions. Evaluated across 11 task categories and 26 benchmarks, the method achieves state-of-the-art performance on diverse modalities (natural, medical, and satellite imagery), significantly improving perceptual quality (LPIPS ↓) and fidelity (NIQE ↓, FID ↓).

0 citationsRead paper