Neurocomputing

UGD-IML: A Unified Generative Diffusion-based Framework for Constrained and Unconstrained Image Manipulation Localization

Yachun Mi1, Xingyang He1, Shixin Sun1, Yu Li1, Zhixuan Li2, Jian Jin3, Chen Hui3, Shaohui Liu1†

1Harbin Institute of Technology

2Nanyang Technological University

3Nanjing University of Information Science and Technology

†Corresponding author

Paper BibTeX

Abstract

Overview

Image manipulation localization must handle both open-world forgeries and constrained settings where an original reference image is available. UGD-IML unifies these tasks with a generative diffusion framework that iteratively denoises a manipulation mask. A conditional control module extracts image guidance, while task-specific conditioning and noise embeddings allow one model to support both settings. Multi-step generation also exposes uncertainty around difficult boundaries and improves robustness across manipulation types.

01 · Figure

UGD-IML formulates manipulation localization as conditional mask generation, using a shared diffusion model for both unconstrained localization and class-constrained localization.

Conditional diffusion architecture of UGD-IML.
UGD-IML formulates manipulation localization as conditional mask generation, using a shared diffusion model for both unconstrained localization and class-constrained localization.

02 · Figure

Unified Task Conditioning

Different conditional-control processes for IML and CIML.
The same conditional control module processes forgery evidence in both tasks. Constrained localization additionally compares the manipulated image with its source image through shared feature extraction.

03 · Figure

Localization Results

UGD-IML qualitative results for constrained and unconstrained manipulation localization.
UGD-IML localizes diverse edited regions with coherent masks in both unconstrained and source-guided scenarios.

04 · Figure

Uncertainty Awareness

Uncertainty and error maps produced by UGD-IML.
Variation across diffusion samples provides an uncertainty estimate that concentrates near ambiguous manipulation boundaries and often aligns with localization errors.

Citation

BibTeX

@article{mi2026ugdiml,
  title={{UGD-IML}: A Unified Generative Diffusion-based Framework for Constrained and Unconstrained Image Manipulation Localization},
  author={Mi, Yachun and He, Xingyang and Sun, Shixin and Li, Yu and Li, Zhixuan and Jin, Jian and Hui, Chen and Liu, Shaohui},
  journal={Neurocomputing},
  year={2026},
  publisher={Elsevier}
}