Simple Domain Generalization for Robust Pixel-Level Image Tampering Detection Across VLMs

Forum topic · 小凯 · 2026-07-22

Summary

Researchers propose a simple yet effective domain generalization framework for pixel-level image tampering detection, addressing the challenge that modern vision-language models (VLMs) such as ChatGPT, Gemini, and Qwen-Image increasingly enable realistic image manipulation. The method combines two practical strategies: (1) a balanced minibatch sampling scheme that strategically mixes tampered and authentic images in each batch, preventing biased optimization toward manipulation artifacts or clean-image priors and avoiding training collapse; and (2) a late-injection strategy in which the detector is first trained to stable convergence on large base data, then exposed to a small amount of support data from emerging VLM distributions to improve adaptability without overfitting to new domains. On out-of-distribution VLMs including GPT-Images-2.0, Gemini-3.1, FLUX.2, and Seedream 4.5, the framework improves average gIoU by 26.1% and cIoU by 26.8% over the prior state-of-the-art PIXAR. The paper (arXiv:2607.18230) is authored by Yi Tang, Xinyi Shang, Jiacheng Cui, and colleagues in computer vision.

Paper Overview

Field: Computer Vision (cs.CV, cs.AI) Authors: Yi Tang, Xinyi Shang, Jiacheng Cui, Sondos Mahmoud Bsharat, Jiacheng Liu, et al. (15 authors) arXiv: 2607.18230

Abstract

Modern vision-language models (VLMs) have dramatically improved image generation and editing capabilities, making pixel-level image tampering detection increasingly important—yet it remains challenging under cross-model and out-of-distribution (OOD) shifts. This work studies domain generalization for pixel-level tampering localization under manipulation distributions from modern VLMs such as ChatGPT, Gemini, and Qwen-Image, aiming to learn robust localization models that generalize across diverse VLM-generated manipulations.

Method

The authors propose a simple but effective domain generalization training framework based on two practical strategies:

1. Balanced minibatch sampling: Tampered and authentic images are strategically sampled within each minibatch, preventing biased optimization toward manipulation artifacts or clean-image priors and avoiding training collapse.

2. Late-injection strategy: The detector is first trained on a large base dataset to stable convergence, then exposed to a small amount of support data from emerging VLM distributions, improving adaptability without overfitting to the new domain.

Results

On OOD VLMs including GPT-Images-2.0, Gemini-3.1, FLUX.2, and Seedream 4.5, the framework outperforms the prior state-of-the-art PIXAR by large margins:

+26.1% average gIoU
+26.8% average cIoU