All Patches Matter, More Patches Better: Enhance AI-Generated Image Detection via Panoptic Patch Learning

2025-04-02

AI Summary

This paper addresses the Few-Patch Bias problem in AI-generated image (AIGI) detection—where existing methods over-rely on a few discriminative patches while neglecting global synthesis artifacts, leading to poor robustness and cross-model generalization. We propose a panoramic patch learning paradigm, pioneering the hypothesis that *all* image patches contain discriminative synthetic traces. To enforce equitable utilization of features across all patches, we introduce two key mechanisms: random patch replacement and patch-level contrastive learning. Our approach fundamentally mitigates local bias induced by lazy learning. Extensive experiments across multiple benchmarks and under both intra-source and cross-source settings demonstrate significant improvements in detection accuracy, alongside reduced overfitting and diminished reliance on localized cues. The code and pretrained models are publicly available.

Technology Category

• Computer Vision: Generative Adversarial Networks (GANs) for Vision
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Machine Learning: Multimodal Learning

Application Category

• Search and Retrieval-Augmented AI: Agentic search
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

The exponential growth of AI-generated images (AIGIs) underscores the urgent need for robust and generalizable detection methods. In this paper, we establish two key principles for AIGI detection through systematic analysis: extbf{(1) All Patches Matter:} Unlike conventional image classification where discriminative features concentrate on object-centric regions, each patch in AIGIs inherently contains synthetic artifacts due to the uniform generation process, suggesting that every patch serves as an important artifact source for detection. extbf{(2) More Patches Better}: Leveraging distributed artifacts across more patches improves detection robustness by capturing complementary forensic evidence and reducing over-reliance on specific patches, thereby enhancing robustness and generalization. However, our counterfactual analysis reveals an undesirable phenomenon: naively trained detectors often exhibit a extbf{Few-Patch Bias}, discriminating between real and synthetic images based on minority patches. We identify extbf{Lazy Learner} as the root cause: detectors preferentially learn conspicuous artifacts in limited patches while neglecting broader artifact distributions. To address this bias, we propose the extbf{P}anoptic extbf{P}atch extbf{L}earning (PPL) framework, involving: (1) Random Patch Replacement that randomly substitutes synthetic patches with real counterparts to compel models to identify artifacts in underutilized regions, encouraging the broader use of more patches; (2) Patch-wise Contrastive Learning that enforces consistent discriminative capability across all patches, ensuring uniform utilization of all patches. Extensive experiments across two different settings on several benchmarks verify the effectiveness of our approach.

Problem

• Detect AI-generated images using all patches for artifacts
• Improve detection robustness by leveraging more patches
• Address Few-Patch Bias with Panoptic Patch Learning

Innovation

• Panoptic Patch Learning framework
• Random Patch Replacement technique
• Patch-wise Contrastive Learning method