Source-linked AI summary
A Sanity Check for AI-generated Image Detection
Shilin Yan, Ouxiang Li, Jiayin Cai, Yanbin Hao, Xiaolong Jiang, Yao Hu, Weidi Xie
TL;DR
The paper asks whether AI-generated image detection is solved despite limited training and evaluation assumptions. It introduces Chameleon, a human-challenging benchmark, and AIDE, which combines low-level patch statistics with high-level semantics. AIDE reaches state-of-the-art results on existing benchmarks, but performance remains low on Chameleon, showing that detection is far from solved.
Problem
Existing detection settings rely on single-generator training and simple evaluation images, providing limited evidence of robustness to realistic, diverse generative outputs.
Method
The paper introduces Chameleon for human-challenging evaluation and AIDE, which combines low-level patch statistics with high-level semantic features.
Results
AIDE achieves state-of-the-art results on existing AIGCDetectBenchmark and GenImage benchmarks, while its performance on Chameleon remains low.
Takeaways & Limitations
AI-generated image detection remains far from solved when evaluated on images that challenge human perception.
Takeaways & Limitations
AIDE's performance remains low on Chameleon, especially for discovering AI-generated images.
Abstract
from arXiv · showhide
With the rapid development of generative models, discerning AI-generated content has evoked increasing attention from both industry and academia. In this paper, we conduct a sanity check on "whether the task of AI-generated image detection has been solved". To start with, we present Chameleon dataset, consisting AIgenerated images that are genuinely challenging for human perception. To quantify the generalization of existing methods, we evaluate 9 off-the-shelf AI-generated image detectors on Chameleon dataset. Upon analysis, almost all models classify AI-generated images as real ones. Later, we propose AIDE (AI-generated Image DEtector with Hybrid Features), which leverages multiple experts to simultaneously extract visual artifacts and noise patterns. Specifically, to capture the high-level semantics, we utilize CLIP to compute the visual embedding. This effectively enables the model to discern AI-generated images based on semantics or contextual information; Secondly, we select the highest frequency patches and the lowest frequency patches in the image, and compute the low-level patchwise features, aiming to detect AI-generated images by low-level artifacts, for example, noise pattern, anti-aliasing, etc. While evaluating on existing benchmarks, for example, AIGCDetectBenchmark and GenImage, AIDE achieves +3.5% and +4.6% improvements to state-of-the-art methods, and on our proposed challenging Chameleon benchmarks, it also achieves the promising results, despite this problem for detecting AI-generated images is far from being solved.
1 INTRODUCTION
Existing AI-generated image detectors are evaluated under limited settings and struggle on images that challenge human perception. The paper introduces Chameleon and AIDE to test and improve robustness across generative models.
- Existing detection research commonly trains on images from a single generative model, limiting exposure to diverse generator properties.
- Chameleon contains AI-generated images that human annotators misclassify as real, spanning diverse categories and high resolutions up to 4K.
- Nine off-the-shelf detectors suffer significant performance drops on Chameleon, frequently classifying challenging AI-generated images as real.
- AIDE combines low-level pixel statistics and noise patterns with high-level semantic features from OpenCLIP.
- AIDE achieves state-of-the-art results on AIGCDetectBenchmark and GenImage while serving as a competitive baseline on Chameleon.
2 RELATED WORKS
Prior AI-generated image detection methods use spatial, frequency, and semantic cues, but their generalization is challenged by evolving generators. Related datasets likewise reflect earlier generator distributions and testing setups.
- Early detectors use spatial-domain cues including color, saturation, co-occurrence, and reflections, but often generalize poorly as generators advance.
- CNNSpot reports generalization from ProGAN-trained classifiers to unseen GAN architectures with careful preprocessing and augmentation.
- FreDect identifies frequency-domain artifacts in GAN-generated images associated with upsampling operations.
- UnivFD uses pretrained CLIP-ViT features, while DIRE introduces reconstruction-based DIRE features for detection.
- Earlier datasets primarily collect GAN-generated fakes, including ProGAN training data and GAN-based testing data.
3 CHAMELEON DATASET
The paper defines broader train-test settings and constructs Chameleon as a large, high-quality benchmark of deceptive, diverse images. Its collection and comparison procedures target more realistic evaluation than existing benchmarks.
- 3.1 PROBLEM FORMULATION: The detection task maps an RGB image I to a binary label indicating whether it is AI-generated or camera-captured.
- 3.1 PROBLEM FORMULATION: Existing settings train on one generator family and test across generators, but use simple random images and constrain training diversity.
- 3.2 CHAMELEON DATASET: The dataset includes over 11k high-fidelity AI-generated images and a similar scale of real photographs collected using identical query words.
- 3.1 PROBLEM FORMULATION: The proposed alternative trains on a wide spectrum of generative models and tests on images genuinely challenging for human perception.
- 3.2 CHAMELEON DATASET: Chameleon is designed around deceptive realism, diverse categories, high image quality, and documented usage permissions.
- 3.2 CHAMELEON DATASET: Chameleon contains approximately 26,000 test images with resolutions from 720P to 4K, providing a larger and more demanding evaluation benchmark.
4 METHODOLOGY
AIDE combines semantic embeddings with frequency-selected patch features to distinguish AI-generated images from photographed images. Its pipeline scores patches by DCT frequency, extracts noise-pattern features from extreme patches, and fuses them for discrimination.
- AIDE Overview: AIDE combines low-level patch statistics, high-level semantic embeddings, and a discriminator for AI-generated image classification.The model targets distinctions between generated and camera-captured images.
- Patch Selection via DCT Scoring: The method divides RGB images into 32 × 32 patches and applies the discrete cosine transform to represent their frequency-domain content.Each patch is represented as an N × N × 3 tensor with N = 32 pixels.
- Patch Selection via DCT Scoring: AIDE scores patches with K band-pass filters over logarithmic absolute DCT coefficients, then sorts the grades to select frequency extremes.The scoring mechanism sums filtered frequency responses across patch positions.
- Patchwise Feature Encoder: It selects the top k highest-frequency and lowest-frequency patches, resizes them to 256 × 256 pixels, and processes them with SRM and two ResNet-50 encoders.The two encoders produce separate highest- and lowest-frequency feature sets before mean pooling.
- Semantic Feature Embedding: For high-level semantics, AIDE uses ConvNeXt-based OpenCLIP features, a linear projection, and mean spatial pooling to capture global context.The resulting semantic representation is combined with patchwise features for discrimination.
- Discriminator: The discriminator channel-wise concatenates low-level and semantic representations, then uses an MLP containing a linear layer, GELU, and classifier to produce the score.This mixture-expert design combines complementary feature representations for final classification.
5 EXPERIMENTS
Experiments evaluate detector generalization on standard benchmarks and the challenging Chameleon set, while ablations examine AIDE’s semantic and frequency-based components. AIDE improves over prior methods on available benchmarks, but all methods degrade sharply on Chameleon and under perturbations.
- Benchmark Results: AIDE achieves a 4.6% improvement in average accuracy over PatchCraft on GenImage.Across eight contemporary generators, AIDE achieves an average accuracy of 86.88%.
- Chameleon Results: Most detectors perform close to random guessing, about 50% accuracy, on the Chameleon benchmark.PatchCraft and other methods experience substantial performance declines despite strong results on existing benchmarks.
- Chameleon Results: AIDE’s performance on Chameleon remains low, especially for discovering AI-generated images.This indicates that the human-challenging Chameleon dataset remains difficult for AI detectors.
- Robustness: JPEG compression and Gaussian blur reduce detector performance by disrupting pixel distributions and diminishing generative artifacts.The robustness evaluation averages accuracy over 16 perturbed B1 test sets.
- Ablation Studies: Removing either high- or low-frequency patches degrades accuracy because the branches capture smooth high-frequency regions and underlying noise, respectively.The patchwise feature extraction ablation evaluates the roles of both frequency ranges.
- Ablation Studies: Removing the semantic branch lowers performance from 92.77% to 76.70%.Without semantic features, AIDE struggles to detect images containing semantic artifacts.
- Ablation Studies: Visualizations show that semantic and patchwise features detect complementary AI-generated cues.Images with semantic errors can evade the semantic-branch ablation, while images without semantic errors can retain unrecognized low-level differences without patchwise extraction.
6 CONCLUSION
The paper argues that AI-generated image detection remains unsolved despite AIDE’s state-of-the-art performance on existing and Chameleon benchmarks.
- The study finds that current detection settings have unreasonable assumptions and that 9 off-the-shelf detectors suffer significant performance declines on Chameleon.Chameleon is manually designed to challenge human perception.
- AIDE combines low-level patch statistics with high-level semantics for AI-generated image detection.The architecture is described as simple yet effective.
- AIDE achieves state-of-the-art performance on AIGCDetectBenchmark, GenImage, and Chameleon compared with previous detectors.The authors state that significant room for future improvement remains.
A EXPERIMENTAL DETAILS
The evaluated baselines represent frequency-, gradient-, semantic-, reconstruction-, and feature-fusion approaches to AI-generated image detection.
- The baselines span frequency-based, gradient-based, semantic-based, reconstruction-based, and other representative detection methods.This selection is intended for comparison with AIDE.
- CNNSpot uses a simple image classifier with JPEG compression and Gaussian blur augmentation to generalize across unknown GAN architectures.
- FreDect classifies GAN-generated images using artifacts observed in the frequency domain.
- Fusing combines global spatial information and local informative features through a two-branch classifier.
- LNP extracts image noise patterns with a learnable denoising network, while LGrad uses gradients from a pretrained CNN to represent generalized artifacts.
- UnivFD trains a binary linear classifier on CLIP features, whereas DIRE uses discrepancies between images and their diffusion-model reconstructions.
- PatchCraft compares rich-texture and poor-texture patches to extract inter-pixel correlation discrepancies as a classification fingerprint.
- NPR studies generalized forgery artifacts produced by CNN generator upsampling beyond frequency-based artifacts.
A.2 STATISTICS OF PUBLIC BENCHMARKS
The public benchmarks use generator-specific training data and evaluate detection across generated-image sources.
- AIGCDetectBenchmark is trained on ProGAN and tested on 16 sets containing images from GAN and Stable Diffusion models.
- GenImage is trained on Stable Diffusion V1.4 and tested on data generated by Stable Diffusion models.
B.1 AP RESULT
The appendix reports AP-focused evaluation and analyzes patch and backbone choices, including a substantial accuracy difference between CLIP-ConvNeXt and supervised CLIP-ViT.
- AP RESULT: Average Precision evaluates performance across decision thresholds and is particularly relevant for imbalanced datasets.
- AP RESULT: AIDE achieves state-of-the-art AP performance among the reported baselines.The authors frame this as evidence of effectiveness in distinguishing classes across thresholds.
- PATCH NUMBER AND PATCH SIZE: Patch size affects detection because overly large patches add irrelevant information, whereas overly small patches may lack sufficient information.
- PATCH NUMBER AND PATCH SIZE: Patch number and patch size are correlated in the ablation analysis.
- CONVNEXT AND VIT: 92.77% average accuracy for CLIP-ConvNeXt exceeds the 80.87% achieved by CLIP-ViT on Benchmark 1.The authors speculate that CNN-supervised architectures learn more low-level information, which is crucial for highly realistic images.
C LIMITATIONS
The method performs competitively on public datasets and Chameleon, but remains unsatisfactory on its own dataset, whose limited scale also constrains future advancement. Chameleon's human-deceptive images create misuse risks requiring controlled access.
- Performance on Chameleon remains unsatisfactory despite state-of-the-art results on public datasets and highly competitive performance there.
- Chameleon’s limited scale leaves room for expansion to better facilitate progress in AI-generated image detection.
- Chameleon’s ability to surpass the human “Turing Test” creates a risk of fictitious social-media profiles and misinformation.
- Dataset access requires a reviewed and approved signed EULA to support compliance with ethical usage protocols.