GAMMA: Generalizable Alignment via Multi-task and Manipulation-Augmented Training for AI-Generated Image Detection

TL;DR

Detectors for AI-generated images often fail to generalize to unseen models due to generation-specific artifacts. GAMMA introduces manipulation augmentation (semantic-aligned inpainting and model-agnostic copy-move/splicing), dual segmentation heads for pixel-level attribution, and a reverse cross-attention mechanism that lets segmentation guide classification, forming a unified framework for robust generalization. It achieves a 5.8 percentage-point improvement over prior best methods on GenImage and shows strong robustness across GPT-4o-based and other diverse datasets, validated by comprehensive ablations. The approach significantly reduces domain bias and enhances semantic reliability, offering a transferable solution for detecting synthetic imagery across evolving generative models.

Abstract

With generative models becoming increasingly sophisticated and diverse, detecting AI-generated images has become increasingly challenging. While existing AI-genereted Image detectors achieve promising performance on in-distribution generated images, their generalization to unseen generative models remains limited. This limitation is largely attributed to their reliance on generation-specific artifacts, such as stylistic priors and compression patterns. To address these limitations, we propose GAMMA, a novel training framework designed to reduce domain bias and enhance semantic alignment. GAMMA introduces diverse manipulation strategies, such as inpainting-based manipulation and semantics-preserving perturbations, to ensure consistency between manipulated and authentic content. We employ multi-task supervision with dual segmentation heads and a classification head, enabling pixel-level source attribution across diverse generative domains. In addition, a reverse cross-attention mechanism is introduced to allow the segmentation heads to guide and correct biased representations in the classification branch. Our method achieves state-of-the-art generalization performance on the GenImage benchmark, imporving accuracy by 5.8%, but also maintains strong robustness on newly released generative model such as GPT-4o.

Figures (3)

• Figure 1: The Overview of GAMMA. Given an input image $\mathbf{I}\in\mathbb{R}^{H\times W\times 3}$, multi-scale features $\mathbf{F}_i$ are extracted via the MiT-B2 encoder and unified through MLP projection and upsampling. Two parallel decoders with identical architectures yield decoder outputs $\mathbf{F}_{\text{Dec}}^{AI}$ and $\mathbf{F}_{\text{Dec}}^{Ma}$, where $\mathbf{F}_{\text{Dec}} \in\mathbb{R}^{\frac{H}{4}\times\frac{W}{4}\times C_{\text{Dec}}}$; while the classifier predicts the global authenticity label with the aid of Reverse Cross-Attention, enabling the segmentation branches to rectify potentially biased classification responses (Fig. \ref{['fig:crossattn']}).
• Figure 2: Different manipulation categories. In inpainting, unmasked regions also pass through diffusion and induce subtle pixel changes, yielding an all-white Mask_AI.
• Figure 3: Impact of different Cross Attention (CA.) configurations on segmentation (AI Seg.) and classification performance. Cls. Prob. denotes the predicted probability of the image being AI-generated.