ICML 2026 · Universal AIGI Detection

OmniAID

Decoupling Semantics and Artifacts for Universal AI-Generated Image Detection in the Wild

Yuncheng Guo¹ · Junyan Ye^2,1 · Chenjue Zhang³ · Hengrui Kang^4,1 · Haohuan Fu³ · Conghui He¹ · Weijia Li^3,1*

¹Shanghai AI Laboratory · ²Sun Yat-sen University · ³Tsinghua SIGS · ⁴Shanghai Jiao Tong University

🚀 Try Online Demo 📄 Paper 💻 Code 🤗 Models 🗂️ Dataset

Core idea

OmniAID decouples semantic flaws and generator artifacts

Decouple what is generated from how it is generated: semantic experts handle content-specific flaws, while the artifact expert remains active for universal forensic traces.

Why OmniAID?

Universal detection needs both semantic reasoning and artifact evidence.

Previous detectors often collapse under semantic shift because they learn one entangled feature space. OmniAID explicitly separates what is generated from how it is generated, then routes each image to the most relevant evidence.

Decoupled MoEsemantic experts + always-active artifact expert

Routable evidenceinterpretable expert routing for mixed or ambiguous content

Mirage benchmarkmodern in-the-wild data for high-fidelity generators

97.2%GenImage accuracyOmniAID-Mirage

91.4%Chameleon accuracyin-the-wild benchmark

88.4%Mirage-Test accuracymodern public test set

Router visualization showing semantic expert selection — Interpretable routing: semantic experts are selected per input, while the artifact expert stays active.

Abstract

Universal AIGI detection by explicit decoupling

Modern image generators are increasingly photorealistic, while existing detectors often learn a single entangled representation that mixes content-dependent semantic flaws with content-agnostic generation artifacts.

OmniAID introduces a decoupled Mixture-of-Experts architecture with routable semantic experts and an always-active universal artifact expert, trained using a two-stage strategy for robust in-the-wild AIGI detection.

Motivation

Semantic gaps and outdated benchmarks limit previous detectors

UnivFD semantic generalization heatmap — UnivFD cross-domain gaps

Effort semantic generalization heatmap — Effort cross-domain gaps

GenImage to Chameleon collapse — Collapse on real-world data

Method

Two-stage decoupled training with a hybrid orthogonal MoE

OmniAID architecture overview and two-stage training — Main pipeline: experts are first specialized with targeted evidence, then a lightweight router integrates semantic experts with the universal artifact expert.

Expert specialization

Semantic experts focus on domain-specific logic flaws, while the artifact expert learns content-agnostic reconstruction traces from aligned real/fake pairs.

Router integration

A global router selects the most relevant semantic experts for each input. The artifact expert is always active to provide universal forensic evidence.

SVD-based orthogonal MoE weight composition — SVD-based orthogonal residual subspace for sparse MoE adaptation.

Results

Robust in-the-wild detection without real/fake bias collapse

+29.3accuracy points over Effort on Chameleon

Balancedhigh real and fake accuracy, instead of detecting only one class

ModernOmniAID-Mirage is trained on contemporary in-the-wild data

Feature analysis

Decoupled feature space and interpretable routing

Effort t-SNE — Effort: entangled baseline representation

OmniAID t-SNE — OmniAID: clearer semantic and real/fake structure

Dataset

Mirage: a modern benchmark for in-the-wild threats

✓ Modern generators up to 2025

✓ In-the-wild synthetic images

✓ Semantic categories: Human, Animal, Object, Scene, Anime

✓ Held-out high-fidelity generators for testing

Citation

@article{guo2025omniaid,
  title={OmniAID: Decoupling Semantics and Artifacts for Universal AI-Generated Image Detection in the Wild},
  author={Guo, Yuncheng and Ye, Junyan and Zhang, Chenjue and Kang, Hengrui and Fu, Haohuan and He, Conghui and Li, Weijia},
  journal={arXiv preprint arXiv:2511.08423},
  year={2025}
}