Awesome AI-Generated Image Detection

A curated and comprehensive list of papers, datasets, and resources for AI-Generated Image Detection (AIGC Detection). This collection covers methods for detecting images synthesized by GANs, diffusion models, autoregressive models, and other generative approaches.

Continuously updated.

Last verified: 2026-03-16 | Total papers: 170+ | Inclusion criteria: Peer-reviewed papers (CVPR, ICCV, ECCV, NeurIPS, ICML, ICLR, AAAI) and high-impact arXiv preprints on AI-generated image detection. Preference for methods providing open-source code.

---

Table of Contents

• Survey
• Paper Tree
• SOTA Methods
• Recommended Benchmarks
• Papers by Venue
  • CVPR 2026
  • ICLR 2026
  • AAAI 2026
  • ICCV 2025
  • ICML 2025
  • CVPR 2025
  • ICLR 2025
  • NeurIPS 2025
  • ECCV 2024
  • NeurIPS 2024
  • CVPR 2024
  • ICML 2024
  • Earlier Works (2020-2023)
• Papers by Category
  • 1. CLIP / Vision-Language Methods
  • 2. Reconstruction-based Methods
  • 3. Frequency-domain Methods
  • 4. Patch / Texture-based Methods
  • 5. Low-level Artifact Detection
  • 6. Perturbation / Robustness-based Methods
  • 7. LMM / Reasoning-based Methods
  • 8. Zero-shot & Training-free Methods
  • 9. Diffusion-specific Detection
  • 10. Autoregressive Model Detection
  • 11. Deepfake / Face Forgery Detection
  • 12. Generalization: Training Strategy & Data Engineering
  • 13. Image Attribution / Source Tracing
• Related Resources

---

Survey

• A Survey on Detecting AI-Generated Images: Key Techniques and Future Directions | arXiv 2025
• A Survey on AI-Generated Image Detection: Techniques, Datasets, and Future Directions | arXiv 2024
• Deepfake Generation and Detection: A Benchmark and Survey | arXiv 2024

---

Paper Tree

Awesome AIGC Image Detection
│
├── 1. CLIP / Vision-Language Methods
│   ├── 1.1 CLIP Fine-tuning (CLIPping, Raising the Bar, DeeCLIP, ForgeLens)
│   ├── 1.2 Prompt & Language-Guided (AntifakePrompt, PLOT, FatFormer, IAPL)
│   ├── 1.3 Feature Decoupling & Fusion (NS-Net, CausalCLIP, CLIPMoLE, CO-SPY)
│   └── 1.4 Information Bottleneck (VIB, MCIB)
│
├── 2. Reconstruction-based Methods
│   ├── 2.1 Diffusion Reconstruction (DIRE, DRCT, LATTE)
│   ├── 2.2 Autoencoder Reconstruction (AEROBLADE, GRRE, CINEMAE)
│   ├── 2.3 Latent Space (LaRE², FakeInversion)
│   └── 2.4 Semantic-Aware (SARE, Exposing the Fake)
│
├── 3. Frequency-domain Methods
│   ├── Spectral Analysis (Fractal Self-Similarity, SPAI, Spectral Artifacts)
│   ├── Color / Demosaicing (Secret Lies in Color, Color Matters)
│   └── Fourier / DCT (Fourier Spectrum Discrepancies)
│
├── 4. Patch / Texture-based Methods
│   ├── Patch Learning (SSP, PatchCraft, Panoptic Patch)
│   └── Texture Analysis (TextureCrop, Breaking Semantic Artifacts)
│
├── 5. Low-level Artifact Detection
│   ├── Up-sampling Artifact (NPR)
│   ├── Pixel-level (FerretNet, PiD, Pixel-level Mapping)
│   ├── Entropy / Bit-plane (MLEP, LOTA)
│   └── Noise / Fingerprint (Implicit Noise Imprint, FingerprintNet)
│
├── 6. Perturbation / Robustness-based Methods
│   ├── Perturbation (RIGID, RA-Det, DEnD)
│   ├── Uncertainty (WePe)
│   └── Distribution Fitting (ConV)
│
├── 7. LMM / Reasoning-based Methods
│   ├── Grounding & Explanation (LEGION, AIGI-Holmes, FakeReasoning)
│   ├── Multi-modal Detection (ThinkFake, Spot the Fake, SIDA)
│   └── VLM-based (Veritas, FakeShield, FAKEXPLAIN)
│
├── 8. Zero-shot & Training-free Methods
│   ├── Training-free (RIGID, WaRPAD, HFI, AEROBLADE, ConV)
│   └── Zero-shot (Forensic Self-Descriptions, Cozzolino, Beyond Generation)
│
├── 9. Diffusion-specific Detection
│   ├── Reconstruction Error (DIRE, AEROBLADE, LaRE², FIRE, DRCT, FakeInversion)
│   └── Probabilistic / Trajectory (Diffusion Epistemic Uncertainty, Corvi 2022)
│
├── 10. Autoregressive Model Detection (PRADA, D³QE, Data Provenance)
│
├── 11. Deepfake Detection (D³, Veritas, Real-Time Deepfake)
│
├── 12. Generalization: Training Strategy & Data Engineering
│   ├── Data Alignment (Dual Alignment, AlignedForensics, B-Free)
│   ├── Real-Centric & Calibration (MIRROR, Stay-Positive, Calibrated)
│   └── Multi-Generator (Community Forensics, GAPL, Forensic-MoE)
│
└── 13. Image Attribution / Source Tracing
    ├── Model Attribution (LatentTracer, AEDR, OCC-CLIP)
    └── AR Attribution (PRADA, Data Provenance)

---

SOTA Methods

Representative methods with open-source code:

---

Recommended Benchmarks

Widely-used benchmarks and evaluation frameworks in this field:

---

Papers by Venue

CVPR 2026

• SimLBR: Learning to Detect Fake Images by Learning to Detect Real Images | Dhakal et al. | Training Strategy
• Scaling Up AI-Generated Image Detection via Generator-Aware Prototypes | Qin et al. | Code | Multi-Generator
• Towards Generalizable AI-Generated Image Detection via Image-Adaptive Prompt Learning | Li et al. | Code | Prompt Learning
• Layer Consistency Matters: Elegant Latent Transition Discrepancy for Generalizable Synthetic Image Detection | Yang et al. | Code | Latent

ICLR 2026

• Unveiling Perceptual Artifacts: A Fine-Grained Benchmark for Interpretable AI-Generated Image Detection (X-AIGD) | Xiao et al. | Code | Benchmark
• FakeXplain: AI-Generated Image Detection via Human-Aligned Grounded Reasoning | Ji et al. | LMM
• Semantic Visual Anomaly Detection and Reasoning in AI-Generated Images | Tan et al. | LMM
• No Pixel Left Behind: A Detail-Preserving Architecture for Robust High-Resolution AI-Generated Image Detection | Mu et al. | Patch
• All Patches Matter, More Patches Better: Enhance AI-Generated Image Detection via Panoptic Patch Learning | Yang et al. | Patch
• HSIC Bottleneck for Cross-Generator and Domain-Incremental Synthetic Image Detection | Yang et al. | Multi-Generator
• Enabling Your Forensic Detector Know How Well It Performs on Distorted Samples | Li et al. | Robustness
• Data Provenance for Image Auto-Regressive Generation | Zhao et al. | AR / Attribution
• A Rich Knowledge Space for Scalable Deepfake Detection | Jung et al. | Deepfake
• Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning | Tan et al. | Oral | Code | LMM

AAAI 2026

• AEDR: Training-Free AI-Generated Image Attribution via Autoencoder Double-Reconstruction | Wang et al. | Oral | Attribution
• Your AI-Generated Image Detector Can Secretly Achieve SOTA Accuracy, If Calibrated | Yang et al. | Calibration
• Beyond Semantic Features: Pixel-level Mapping for Generalized AI-Generated Image Detection | Zhou et al. | Pixel-level
• Aggregating Diverse Cue Experts for AI-Generated Image Detection | Tan et al. | Feature Fusion
• Explainable Synthetic Image Detection through Diffusion Timestep Ensembling | Wu et al. | LMM

NeurIPS 2025

• Detecting Generated Images by Fitting Natural Image Distributions | Zhang et al. | Code | Spotlight Training-free
• Denoising Trajectory Biases for Zero-Shot AI-Generated Image Detection | Liang et al. | Training-free
• Training-free Detection of AI-generated images via Cropping Robustness | Choi et al. | Code | Training-free
• Towards Generalizable Detector for Generated Image | Cai et al. | Code | Perturbation
• Epistemic Uncertainty for Generated Image Detection | Nie et al. | Code | Uncertainty
• Dual Data Alignment Makes AI-Generated Image Detector Easier Generalizable | Chen et al. | Spotlight Training Strategy
• MLEP: Multi-granularity Local Entropy Patterns for Generalized AI-generated Image Detection | Yuan et al. | Code | Entropy
• FerretNet: Efficient Synthetic Image Detection via Local Pixel Dependencies | Liang et al. | Code | Pixel
• Breaking Latent Prior Bias in Detectors for Generalizable AIGC Image Detection | Zhou et al. | Bias-Free

ICCV 2025

• LEGION: Learning to Ground and Explain for Synthetic Image Detection | Kang et al. | Code | LMM
• AIGI-Holmes: Towards Explainable and Generalizable AI-Generated Image Detection via Multimodal Large Language Models | Zhou et al. | Code | LMM
• ForgeLens: Data-Efficient Forgery Focus for Generalizable Forgery Image Detection | Chen et al. | Code | CLIP
• LOTA: Bit-Planes Guided AI-Generated Image Detection | Wang et al. | Code | Bit-plane
• Forensic-MoE: Exploring Comprehensive Synthetic Image Detection Traces with Mixture of Experts | Fang et al. | Code | Multi-Generator MoE
• Semantic Discrepancy-aware Detector for Image Forgery Identification | Wang et al. | Reconstruction
• D³QE: Learning Discrete Distribution Discrepancy-aware Quantization Error for Autoregressive-Generated Image Detection | Zhang et al. | Autoregressive
• Bridging the Gap Between Ideal and Real-world Evaluation: Benchmarking AI-Generated Image Detection | Li et al. | Benchmark
• Diffusion Epistemic Uncertainty with Asymmetric Learning for Diffusion-Generated Image Detection | Huang et al. | Diffusion

ICML 2025

• Are High-Quality AI-Generated Images More Difficult for Models to Detect? | Xiao et al. | Code | Analysis
• Few-Shot Learner Generalizes Across AI-Generated Image Detection | - | Code | Few-shot
• Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection | Yan et al. | Code | CLIP
• Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection | - | Code | Training Strategy
• Unlocking the Capabilities of Large Vision-Language Models for Generalizable and Explainable Deepfake Detection | - | LMM
• PiD: Generalized AI-Generated Images Detection with Pixelwise Decomposition Residuals | Fu et al. | Reconstruction

CVPR 2025

• Any-Resolution AI-Generated Image Detection by Spectral Learning | Karageorgiou et al. | Code | Frequency
• A Bias-Free Training Paradigm for More General AI-generated Image Detection | Guillaro et al. | Code | Training Strategy
• Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network | Zhang et al. | Code | CLIP
• FIRE: Robust Detection of Diffusion-Generated Images via Frequency-Guided Reconstruction Error | Chu et al. | Reconstruction
• Secret Lies in Color: Enhancing AI-Generated Images Detection with Color Distribution Analysis | Jia et al. | Color
• Community Forensics: Using Thousands of Generators to Train Fake Image Detectors | Park et al. | Code | Multi-Generator
• SIDA: Social Media Image Deepfake Detection, Localization and Explanation with LMM | Huang et al. | Code | LMM
• CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI | Cheng et al. | Feature Fusion
• D³: Scaling Up Deepfake Detection by Learning from Discrepancy | Yang et al. | Code | Training Strategy
• Forensic Self-Descriptions Are All You Need for Zero-Shot Detection, Open-Set Source Attribution, and Localization | Nguyen et al. | Zero-shot
• Beyond Generation: A Diffusion-based Low-level Feature Extractor for Detecting AI-generated Images | Zhong et al. | Zero-shot

ICLR 2025

• Aligned Datasets Improve Detection of Latent Diffusion-Generated Images | Rajan et al. | Code | Training Strategy
• A Sanity Check for AI-generated Image Detection | - | Code | Benchmark Feature Fusion
• LOKI: A Comprehensive Synthetic Data Detection Benchmark using Large Multimodal Models | - | Code | Benchmark
• FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models | Xu et al. | Code | LMM

ECCV 2024

• Zero-Shot Detection of AI-Generated Images | Cozzolino et al. | Zero-shot
• Contrasting Deepfakes Diffusion via Contrastive Learning and Global-Local Similarities | Baraldi et al. | Contrastive
• Leveraging Representations from Intermediate Encoder-Blocks for Synthetic Image Detection | Koutlis et al. | Code | CLIP

NeurIPS 2024

• Breaking Semantic Artifacts for Generalized AI-generated Image Detection | Zheng et al. | Code | Texture

CVPR 2024

• FakeInversion: Learning to Detect Images from Unseen Text-to-Image Models by Inverting Stable Diffusion | Cazenavette et al. | Page | Reconstruction
• AEROBLADE: Training-Free Detection of Latent Diffusion Images Using Autoencoder Reconstruction Error | Ricker et al. | Reconstruction
• LaRE²: Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection | Luo et al. | Reconstruction
• MaskSim: Detection of Synthetic Images by Masked Spectrum Similarity Analysis | Li et al. | Frequency
• Forgery-aware Adaptive Transformer for Generalizable Synthetic Image Detection | Liu et al. | Code | Adaptation

ICML 2024

• How to Trace Latent Generative Model Generated Images without Artificial Watermark? | Wang et al. | Code | Attribution
• DRCT: Diffusion Reconstruction Contrastive Training towards Universal Detection | Chen et al. | Code | Reconstruction

Earlier Works (2020-2023)

ICCV 2023

• DIRE for Diffusion-Generated Image Detection | Wang et al. | Code | Reconstruction

CVPR 2023

• Towards Universal Fake Image Detectors that Generalize Across Generative Models | Ojha et al. | Code | CLIP
• Learning on Gradients: Generalized Artifacts Representation for GAN-Generated Images Detection | Tan et al. | Code | Gradient

ECCV 2022

• Detecting Generated Images by Real Images | Tang et al. | Code | Real-Only
• FingerprintNet: Synthesized Fingerprints for Generated Image Detection | - | Fingerprint
• On The Detection of Synthetic Images Generated by Diffusion Models | Corvi et al. | Code | Diffusion

CVPR 2020 & NeurIPS 2020

• CNN-generated images are surprisingly easy to spot...for now | Wang et al. | Code | CNN Forensics
• Watch your Up-Convolution: CNN Based Generative Deep Neural Networks are Failing to Reproduce Spectral Distributions | Durall et al. | Frequency
• Fourier Spectrum Discrepancies in Deep Network Generated Images | - | Code | Frequency
• Leveraging Frequency Analysis for Deep Fake Image Recognition | Frank et al. | Code | Frequency

Other Notable Works (arXiv)

• Detecting Generated Images by Real Images Only | 2023 | Real-Only
• Transfer Learning of Real Image Features with Soft Contrastive Loss for Fake Image Detection | 2024 | Training Strategy
• RIGID: A Training-free and Model-Agnostic Framework for Robust AI-Generated Image Detection | He et al. 2024 | Perturbation
• GenDet: Towards Good Generalizations for AI-Generated Image Detection | 2023 | Training Strategy
• Generalizable Synthetic Image Detection via Language-guided Contrastive Learning (LASTED) | Code | 2023 | CLIP
• AntifakePrompt: Prompt-Tuned Vision-Language Models are Fake Image Detectors | Chang et al. 2023 | Prompt
• PatchCraft: Exploring Texture Patch for Efficient AI-generated Image Detection | Zhong et al. 2024 | Patch
• Bi-LORA: A Vision-Language Approach for Synthetic Image Detection | 2024 | CLIP
• Fake or JPEG? Revealing Common Biases in Generated Image Detection Datasets | Grommelt et al. 2024 | Page | Bias

---

Papers by Category

1. CLIP / Vision-Language Methods

Leverage CLIP or other vision-language models whose pre-trained features already encode subtle generative artifacts, enabling strong cross-generator generalization with minimal fine-tuning.

1.1 CLIP Fine-tuning & Adaptation

Adapt CLIP's visual encoder via linear probes, LoRA, or lightweight adapters to extract forgery-discriminative features with minimal parameter updates.

• Raising the Bar of AI-generated Image Detection with CLIP | Cozzolino et al. | 2024 | CLIP
• CLIPping the Deception: Adapting Vision-Language Models for Universal Deepfake Detection | ICMR 2024 | Code | CLIP
• DeeCLIP: A Robust and Generalizable Transformer-Based Framework | 2025 | CLIP
• DGS-Net: Distillation-Guided Gradient Surgery for CLIP Fine-Tuning in AI-Generated Image Detection | Yan et al. 2025 | CLIP
• NS-Net: Decoupling CLIP Semantic Information through NULL-Space for Generalizable AI-Generated Image Detection | Yan et al. 2025 | CLIP
• When Semantics Regulate: Rethinking Patch Shuffle for AI-Generated Image Detection | 2025 | CLIP
• ForgeLens: Data-Efficient Forgery Focus for Generalizable Forgery Image Detection | Chen et al. | ICCV 2025 | Code | CLIP

1.2 Prompt Learning & Language-Guided

Steer CLIP's representation via learnable prompts or language-guided objectives rather than weight adaptation, enabling few-shot domain transfer.

• AntifakePrompt: Prompt-Tuned Vision-Language Models are Fake Image Detectors | Chang et al. 2024 | Prompt
• Forgery-aware Adaptive Transformer for Generalizable Synthetic Image Detection | Liu et al. | CVPR 2024 | Code | Language-Guided
• MiraGe: Multimodal Discriminative Representation Learning for Generalizable AI-Generated Image Detection | Shi et al. 2025 | Prompt
• Towards Generalizable AI-Generated Image Detection via Image-Adaptive Prompt Learning | Li et al. | CVPR 2026 | Code | Prompt

1.3 Feature Fusion

Decompose or fuse heterogeneous features—semantic, frequency, pixel-level—to separate forgery clues from content and improve cross-domain robustness.

• CausalCLIP: Causally-Informed Feature Disentanglement and Filtering for Generalizable Detection | Liu et al. 2025 | CLIP
• Mixture of Low-rank Experts for Transferable AI-Generated Image Detection | Liu et al. 2024 | Code | CLIP
• Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection | Yan et al. | ICML 2025 | Code | CLIP
• CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI | Cheng et al. | CVPR 2025 | Feature Fusion
• Aggregating Diverse Cue Experts for AI-Generated Image Detection | Tan et al. | AAAI 2026 | Feature Fusion
• AIDE: A Sanity Check for AI-generated Image Detection | Yan et al. | ICLR 2025 | Code | Feature Fusion

1.4 Information Bottleneck

Apply information bottleneck principles to compress away content and retain only detection-relevant signals, improving robustness to distribution shift.

• Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network | Zhang et al. | CVPR 2025 | Code | CLIP
• Multimodal Conditional Information Bottleneck for Generalizable AI-Generated Image Detection | Qin et al. 2025 | Code | CLIP

---

2. Reconstruction-based Methods

Detect AI-generated images by measuring how well a pre-trained generative model can reconstruct them—synthetic images yield lower error than real ones.

2.1 Diffusion Reconstruction

Compute reconstruction errors via DDIM inversion or latent trajectory encoding under a fixed diffusion model.

• DIRE for Diffusion-Generated Image Detection | Wang et al. | ICCV 2023 | Code | Reconstruction
• DRCT: Diffusion Reconstruction Contrastive Training towards Universal Detection | ICML 2024 | Reconstruction
• LATTE: Latent Trajectory Embedding for Diffusion-Generated Image Detection | Vasilcoiu et al. 2025 | Code | Reconstruction
• STD-FD: Spatio-Temporal Distribution Fitting Deviation for AIGC Forgery Identification | Lou et al. | Reconstruction

2.2 Autoencoder Reconstruction

Exploit VAE or MAE encoder–decoder reconstruction error, which is near-zero for in-distribution generated images but elevated for real ones.

• AEROBLADE: Training-Free Detection of Latent Diffusion Images Using Autoencoder Reconstruction Error | Ricker et al. | CVPR 2024 | Reconstruction
• GRRE: Leveraging G-Channel Removed Reconstruction Error for Robust Detection of AI-Generated Images | He et al. 2026 | Reconstruction
• CINEMAE: Leveraging Frozen Masked Autoencoders for Cross-Generator AI Image Detection | Jang et al. 2025 | Reconstruction
• FIRE: Robust Detection of Diffusion-Generated Images via Frequency-Guided Reconstruction Error | Chu et al. | CVPR 2025 | Reconstruction

2.3 Latent Space

Work in the compressed latent space of a diffusion VAE, using inversion residuals as detection signals.

• LaRE²: Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection | Luo et al. | CVPR 2024 | Reconstruction
• FakeInversion: Learning to Detect Images from Unseen Text-to-Image Models by Inverting Stable Diffusion | Cazenavette et al. | CVPR 2024 | Reconstruction

2.4 Semantic-Aware Reconstruction

Condition reconstruction errors on semantic content to better disentangle generation artifacts from content-related variation.

• SARE: Semantic-Aware Reconstruction Error for Generalizable Diffusion-Generated Image Detection | Kang et al. 2025 | arXiv | Reconstruction
• Exposing the Fake: Effective Diffusion-Generated Images Detection | Ma et al. | AdvML-Frontiers@ICML 2023 Workshop | Reconstruction

---

3. Frequency-domain Methods

Exploit spectral artifacts—fractal self-similarity, missing high-frequency fall-off, color demosaicing inconsistencies—left by generative pipelines in the frequency domain.

• Generalizable AI-Generated Image Detection Based on Fractal Self-Similarity in the Spectrum | Xiao et al. 2025 | Frequency
• Any-Resolution AI-Generated Image Detection by Spectral Learning | Karageorgiou et al. | CVPR 2025 | Code | Frequency
• Improving Synthetic Image Detection Towards Generalization: An Image Transformation Perspective | Li et al. 2025 | Frequency
• Training-Free AI-Generated Image Detection via Spectral Artifacts | Frequency
• Secret Lies in Color: Enhancing AI-Generated Images Detection with Color Distribution Analysis | Jia et al. | CVPR 2025 | Color
• Color Matters: Demosaicing-Guided Color Correlation Training for Generalizable AI-Generated Image Detection | Zhong et al. 2026 | Color
• Fourier Spectrum Discrepancies in Deep Network Generated Images | NeurIPS 2020 | Code | Frequency

---

4. Patch / Texture-based Methods

Extract and compare local patch-level texture features—noise residuals, texture contrast, patch-contrastive signals—to expose synthetic origin independent of semantic content.

• A Single Simple Patch is All You Need for AI-generated Image Detection | Chen et al. 2024 | Patch
• PatchCraft: Exploring Texture Patch for Efficient AI-generated Image Detection | Zhong et al. 2024 | Patch
• All Patches Matter, More Patches Better: Enhance AI-Generated Image Detection via Panoptic Patch Learning | Yang et al. | ICLR 2026 | Patch
• TextureCrop: Enhancing Synthetic Image Detection through Texture-based Cropping | Konstantinidou et al. | Texture
• Breaking Semantic Artifacts for Generalized AI-generated Image Detection | Zheng et al. | NeurIPS 2024 | Code | Texture
• No Pixel Left Behind: A Detail-Preserving Architecture for Robust High-Resolution AI-Generated Image Detection | Mu et al. | ICLR 2026 | Pixel
• Rethinking the Use of Vision Transformers for AI-Generated Image Detection | Park et al. 2025 | Architecture

---

5. Low-level Artifact Detection

Detect fine-grained pixel-level artifacts—upsampling patterns, local pixel dependencies, bit-plane anomalies, entropy deviations—that are architecture-agnostic byproducts of the generation pipeline.

• Rethinking the Up-Sampling Operations in CNN-based Generative Network for Generalizable Deepfake Detection | Tan et al. | CVPR 2024 | Code | Up-sampling Artifact
• FerretNet: Efficient Synthetic Image Detection via Local Pixel Dependencies | Liang et al. | NeurIPS 2025 | Code | Pixel
• Beyond Semantic Features: Pixel-level Mapping for Generalized AI-Generated Image Detection | Zhou et al. | AAAI 2026 | Pixel
• PiD: Generalized AI-Generated Images Detection with Pixelwise Decomposition Residuals | Fu et al. | ICML 2025 | Pixel
• MLEP: Multi-granularity Local Entropy Patterns for Generalized AI-generated Image Detection | Yuan et al. | NeurIPS 2025 | Code | Entropy
• LOTA: Bit-Planes Guided AI-Generated Image Detection | Wang et al. | ICCV 2025 | Code | Bit-plane
• Revealing the Implicit Noise-based Imprint of Generative Models | Li et al. 2025 | Fingerprint
• Perceptual Classifiers: Detecting Generative Images using Perceptual Features | Durbha et al. | ICCV 2025 Workshop | Perceptual
• Beyond Generation: A Diffusion-based Low-level Feature Extractor for Detecting AI-generated Images | Zhong et al. | CVPR 2025 | Zero-shot
• Exploring the Collaborative Advantage of Low-level Information on Generalizable AI-Generated Image Detection | Zhou et al. 2025 | Feature Fusion

---

6. Perturbation / Robustness-based Methods

Exploit the asymmetric response of real vs. synthetic images to controlled perturbations—real images are more stable on the natural manifold, while generated images deviate more.

• RIGID: A Training-free and Model-Agnostic Framework for Robust AI-Generated Image Detection | He et al. 2024 | Perturbation
• RA-Det: Towards Universal Detection of AI-Generated Images via Robustness Asymmetry | Wang et al. 2026 | arXiv | Perturbation
• Detecting AI-Generated Images via Distributional Deviations from Real Images | Niu et al. 2026 | Perturbation
• Epistemic Uncertainty for Generated Image Detection (WePe) | Nie et al. | NeurIPS 2025 | Code | Uncertainty
• How Noise Benefits AI-generated Image Detection | Yan et al. 2025 | Perturbation
• Towards Generalizable Detector for Generated Image | Cai et al. | NeurIPS 2025 | Code | Perturbation
• Detecting Generated Images by Fitting Natural Image Distributions | Zhang et al. | NeurIPS 2025 Spotlight | Code | Training-free
• ENABLING YOUR FORENSIC DETECTOR KNOW HOW WELL IT PERFORMS ON DISTORTED SAMPLES | Li et al. | ICLR 2026 | Robustness

---

7. LMM / Reasoning-based Methods

Leverage large multimodal models to provide not just binary detection but also artifact localization, reasoning, and natural-language explanations of why an image is fake.

• LEGION: Learning to Ground and Explain for Synthetic Image Detection | Kang et al. | ICCV 2025 | Code | LMM
• AIGI-Holmes: Towards Explainable and Generalizable AI-Generated Image Detection via Multimodal Large Language Models | Zhou et al. | ICCV 2025 | Code | LMM
• ThinkFake: Reasoning in Multimodal Large Language Models for AI-Generated Image Detection | Huang et al. 2025 | LMM
• FakeReasoning: Towards Generalizable Forgery Detection and Reasoning | Gao et al. 2025 | LMM
• Spot the Fake: Large Multimodal Model-Based Synthetic Image Detection with Artifact Explanation | Wen et al. 2025 | LMM
• Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning | Tan et al. | ICLR 2026 Oral | Code | LMM
• SIDA: Social Media Image Deepfake Detection, Localization and Explanation with LMM | Huang et al. | CVPR 2025 | Code | LMM
• FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models | Xu et al. | ICLR 2025 | Code | LMM
• FAKEXPLAIN: AI-Generated Images Detection via Human-Aligned Grounded Reasoning | Ji et al. | ICLR 2026 | LMM
• Semantic Visual Anomaly Detection and Reasoning in AI-Generated Images | Tan et al. | ICLR 2026 | LMM
• Explainable Synthetic Image Detection through Diffusion Timestep Ensembling | Wu et al. | AAAI 2026 | LMM
• Unlocking the Capabilities of Large Vision-Language Models for Generalizable and Explainable Deepfake Detection | ICML 2025 | LMM

---

8. Training-free & Zero-shot Methods

Require no training on fake data—exploit intrinsic properties of pre-trained models (reconstruction error, feature distributions, perturbation response) to detect without any task-specific optimization, naturally generalizing to unseen generators.

• RIGID: A Training-free and Model-Agnostic Framework for Robust AI-Generated Image Detection | He et al. 2024 | Training-free

• Intermediate Representations Are Strong Training-Free AI-Generated Image Detectors | Training-free

• Training-Free AI-Generated Image Detection via Spectral Artifacts | Training-free

• Forensic Self-Descriptions Are All You Need for Zero-Shot Detection, Open-Set Source Attribution, and Localization | Nguyen et al. | CVPR 2025 | Zero-shot

• Zero-Shot Detection of AI-Generated Images | Cozzolino et al. | ECCV 2024 | Zero-shot

• Detecting Generated Images by Fitting Natural Image Distributions | Zhang et al. | NeurIPS 2025 Spotlight | Code | Training-free

• General and Domain-Specific Zero-shot Detection of Generated Images via Conditional Likelihood | Betser et al. 2025 | Zero-shot

• Training-free Detection of AI-generated images via Cropping Robustness | Choi et al. | Code | Training-free

• Denoising Trajectory Biases for Zero-Shot AI-Generated Image Detection | Liang et al. | Training-free

• Revisiting Reconstruction-based AI-generated Image Detection: A Geometric Perspective | Jiang et al. 2025 | Training-free

• HFI: A Unified Framework for Training-free Detection and Implicit Watermarking of Latent Diffusion Models | Choi et al. 2024 | Training-free

• Understanding and Improving Training-Free AI-Generated Image Detections with Vision Foundation Models | Tsai et al. 2024 | Training-free

---

9. Diffusion-specific Detection

Core mechanism exploits DDIM inversion, VAE reconstruction, or denoising trajectories specific to diffusion models—would not transfer to GANs or AR generators.

• DIRE for Diffusion-Generated Image Detection | Wang et al. | ICCV 2023 | Code | Diffusion
• AEROBLADE: Training-Free Detection of Latent Diffusion Images Using Autoencoder Reconstruction Error | Ricker et al. | CVPR 2024 | Diffusion Training-free
• LaRE²: Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection | Luo et al. | CVPR 2024 | Diffusion
• FakeInversion: Learning to Detect Images from Unseen Text-to-Image Models by Inverting Stable Diffusion | Cazenavette et al. | CVPR 2024 | Page | Diffusion
• DRCT: Diffusion Reconstruction Contrastive Training towards Universal Detection | Chen et al. | ICML 2024 | Code | Diffusion
• On the Detection of Synthetic Images Generated by Diffusion Models | Corvi et al. | ECCV 2022 | Code | Diffusion
• FIRE: Robust Detection of Diffusion-Generated Images via Frequency-Guided Reconstruction Error | Chu et al. | CVPR 2025 | Diffusion
• Diffusion Epistemic Uncertainty with Asymmetric Learning for Diffusion-Generated Image Detection | Huang et al. | ICCV 2025 | Diffusion
• Exposing the Fake: Effective Diffusion-Generated Images Detection | Ma et al. | AdvML-Frontiers@ICML 2023 Workshop | Diffusion

---

10. Autoregressive Model Detection

Target the unique artifacts of token-by-token autoregressive generation—discrete quantization errors and conditional token probability patterns.

• PRADA: Probability-Ratio-Based Attribution and Detection of Autoregressive-Generated Images | Damm et al. 2025 | Autoregressive
• D³QE: Learning Discrete Distribution Discrepancy-aware Quantization Error for Autoregressive-Generated Image Detection | Zhang et al. 2025 | Autoregressive
• Data Provenance for Image Auto-Regressive Generation | Zhao et al. | ICLR 2026 | AR / Attribution

---

11. Deepfake Detection

Face deepfake methods that also generalize to general AI-generated image detection benchmarks.

• D³: Scaling Up Deepfake Detection by Learning from Discrepancy | Yang et al. | CVPR 2025 | Code | Deepfake + General
• Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning | Tan et al. | ICLR 2026 Oral | Code | LMM + Deepfake
• Real-Time Deepfake Detection in the Real-World | Cavia et al. 2024 | Deepfake

---

12. Generalization: Training Strategy & Data Engineering

Address the generalization gap through training data engineering—bias elimination, semantic alignment, multi-generator scaling, and calibration.

12.1 Data Alignment & Construction

Eliminate spurious dataset biases (JPEG quality, resolution, content mismatch) by constructing semantically aligned real–fake training pairs.

• Dual Data Alignment Makes AI-Generated Image Detector Easier Generalizable | Chen et al. | NeurIPS 2025 Spotlight | Data Alignment
• Aligned Datasets Improve Detection of Latent Diffusion-Generated Images | Rajan et al. | ICLR 2025 | Code | Data Alignment
• A Bias-Free Training Paradigm for More General AI-generated Image Detection | Guillaro et al. | CVPR 2025 | Code | Bias-Free
• Breaking Latent Prior Bias in Detectors for Generalizable AIGC Image Detection | Zhou et al. | NeurIPS 2025 | Bias-Free
• Fake or JPEG? Revealing Common Biases in Generated Image Detection Datasets | Grommelt et al. 2024 | Page | Data Alignment
• Task-Model Alignment: A Simple Path to Generalizable AI-Generated Image Detection | Chen et al. 2025 | Data Alignment
• SimLBR: Learning to Detect Fake Images by Learning to Detect Real Images | Dhakal et al. | CVPR 2026 | Regularization

12.2 Real-Centric & Calibration

Focus on the real-image manifold and calibration rather than fake data engineering—classify anything that deviates as synthetic.

• MIRROR: Manifold Ideal Reference ReconstructOR for Generalizable AI-Generated Image Detection | Liu et al. 2026 | arXiv | Code | Real-Centric
• Your AI-Generated Image Detector Can Secretly Achieve SOTA Accuracy, If Calibrated | Yang et al. | AAAI 2026 | Calibration
• Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection | ICML 2025 | Code | Real-Centric
• Towards Generalizable AI-Generated Image Detection via Image-Adaptive Prompt Learning | Li et al. | CVPR 2026 | Code | Test-Time Adaptation

12.3 Multi-Generator

Scale training to thousands of generative models, empirically showing monotonic generalization improvement with generator diversity.

• Community Forensics: Using Thousands of Generators to Train Fake Image Detectors | Park et al. | CVPR 2025 | Code | Multi-Generator
• Scaling Up AI-Generated Image Detection via Generator-Aware Prototypes | Qin et al. | CVPR 2026 | Code | Multi-Generator
• Forensic-MoE: Exploring Comprehensive Synthetic Image Detection Traces with Mixture of Experts | Fang et al. | ICCV 2025 | Code | Multi-Generator
• HSIC Bottleneck for Cross-Generator and Domain-Incremental Synthetic Image Detection | Yang et al. | ICLR 2026 | Multi-Generator

---

13. Image Attribution / Source Tracing

Go beyond binary detection to identify which generative model produced an image, exploiting model-specific reconstruction fingerprints and latent signatures.

• How to Trace Latent Generative Model Generated Images without Artificial Watermark? | Wang et al. | ICML 2024 | Code | Attribution
• AEDR: Training-Free AI-Generated Image Attribution via Autoencoder Double-Reconstruction | Wang et al. | AAAI 2026 Oral | Attribution
• Which Model Generated This Image? A Model-Agnostic Approach for Origin Attribution | Liu et al. 2024 | Code | Attribution
• PRADA: Probability-Ratio-Based Attribution and Detection of Autoregressive-Generated Images | Damm et al. 2025 | Attribution
• Data Provenance for Image Auto-Regressive Generation | Zhao et al. | ICLR 2026 | Attribution

---

Citation

If you find this repository useful, please consider citing:

@misc{awesome-aigc-image-detection,
  title={Awesome AI-Generated Image Detection},
  author={yjtlab},
  year={2026},
  publisher={GitHub},
  howpublished={\url{https://github.com/yjtlab/awesome-aigc-image-detection}}
}

---

Last Updated: 2026-03-16 | Total Papers: 170+

Contributions and suggestions are welcome! Please open an issue or pull request.