Title: Impostor: An Agent-Driven Benchmark for Realistic AIGC Manipulation Localization

Abstract:

The rapid evolution of generative image editing has significantly enhanced the realism and controllability of localized image modifications, thereby creating new hurdles for image manipulation detection and localization (IMDL). Despite these advancements, current IMDL benchmarks suffer from shortcomings in visual fidelity, diversity of manipulations, and coverage of generative models, failing to adequately capture recent trends in image tampering. To overcome these issues, we present Impostor, a premium dataset for AI-edited image manipulation localization comprising 100,000 manipulated images.

Impostor is built using CraftAgent, a closed-loop agent framework designed to automatically produce visually realistic and diverse manipulated images. This framework seamlessly integrates scene perception, editing planning, manipulation execution, quality validation, and iterative reflection. Additionally, Impostor features images created by seven recent AIGC models, covering three distinct manipulation types and multiple manipulated regions, offering a more robust benchmark for AIGC-based IMDL tasks.

We also introduce PhaseAware-Net (PANet), a semantic-forensic framework that employs local phase modeling and semantic-forensic consistency learning to more effectively localize manipulated regions that are semantically plausible but forensically inconsistent. Our extensive experiments demonstrate that Impostor presents substantial difficulties for existing large vision-language models (LVLMs) and specialized IMDL methods. Meanwhile, PANet delivers superior performance on Impostor as well as several public benchmarks.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC