ARTFEED — Contemporary Art Intelligence

COCO-Inpaint Benchmark Targets Inpainting-Based Image Manipulation Detection

ai-technology · 2026-05-18

Researchers have introduced COCO-Inpaint, a benchmark designed to detect and localize inpainting-based image manipulations. While existing Image Manipulation Detection and Localization (IMDL) methods focus on splicing or copy-move forgeries, inpainting benchmarks are scarce. COCO-Inpaint offers three key contributions: 238,302 high-quality inpainted images generated by six state-of-the-art inpainting models, four mask generation strategies with optional text guidance for diverse scenarios, and large-scale semantic diversity. The benchmark highlights intrinsic inconsistencies between inpainted and authentic regions, aiming to advance multimedia authenticity and security.

Key facts

  • COCO-Inpaint is a benchmark for detecting and localizing inpainting-based image manipulations.
  • It addresses the gap in IMDL methods that mainly target splicing or copy-move forgeries.
  • The benchmark includes 238,302 inpainted images generated by six state-of-the-art inpainting models.
  • Four mask generation strategies with optional text guidance enable diverse generation scenarios.
  • The dataset offers large-scale coverage with rich semantic diversity.
  • It highlights intrinsic inconsistencies between inpainted and authentic regions.
  • The work is published on arXiv with ID 2504.18361.
  • The announcement type is replace-cross.

Entities

Institutions

  • arXiv

Sources