AI-generated content (AIGC) has substantially narrowed the visual gap between natural and synthetic images, while existing detectors often rely on generator-specific artifacts and exhibit limited generalization to unseen models. This motivates a fundamental question: which intrinsic physical image characteristics provide stable, discriminative cues for distinguishing natural from AI-generated images? We investigate this by systematically identifying robust physical features and incorporating them into a vision-language representation without external textual prompts at inference. We first analyze 15 handcrafted physical descriptors across images from over 20 generative models. Based on cross-generator stability and real-fake discriminability, we propose a Feature Stability-Discriminability Validation Assessment (FSDVA) to select a compact set of reliable features. Five core features, including Laplacian variance, Sobel statistics, and residual-noise variance, are then used to construct structured textual descriptions encoding physical evidence for CLIP-based learning. We integrate these physics-enriched descriptions with image representations via a native physics-guided multimodal framework, adapting CLIP ViT-L 14 with LoRA. At inference, the model requires only the input image, with no external captions or manual prompts. On GenImage, our method achieves 99.0 percent accuracy on SDv1.4 and 96.7 percent on Wukong, with a mean of 97.23 percent across eight generative models versus 95.8 percent for C2pCLIP under the same setting. On UniversalFakeDetect, it maintains strong, stable generalization across diverse and unseen generation settings. These results show that carefully selected physical image statistics provide complementary forensic evidence to pretrained vision-language representations, enabling interpretable, more generator-agnostic AIGC detection.
Figures & tables
Figure 1: Overall workflow of the proposed synthetic image detection framework. (a) Enhanced caption preparation: Integrate input images’ core physical features into captions to enrich representation. (b) Training: Use enhanced captions and class prompts to train the model for real/AI-generated image discrimination. (c) Testing: Input images into the trained model to predict AI generation. Our training pipeline framework natively supports end-to-end training without additional architectural adjustments.
F1=HW1∑i,jHSVi,j,1
F2=HW1∑i,jHSVi,j,2
F3=log(1+Var(∇2Ig))
F4=10⋅HW1∑i,jSx2+Sy2
F5=10⋅HW1∑i,j(Mi,j−μM)2
F6=log(1+N1∑kVar(DCT(Bk)))
F7=log(1+1000Var(R))
F8=−∑bpblog2pb
F9=Var(R)Var(G)Cov(R,G)
F10=Var(R)Var(B)Cov(R,B)
F11=−∑vpvlog2pv
F12=−∑vpvlog2pv
F13=log(1+Var(HSV:,:,0))
F14=HW1∑i,j255Ei,j
F15=log(1+∣Kurt(B)∣)
Table 1: Mathematical definitions of 15 synthetic image detection features. Key symbol definitions: i,j = 2D pixel coordinates (height/width); H,W = image height/width; HSVi,j,c = HSV color space value at pixel (i,j) for channel c (1=saturation, 2=brightness, 0=hue); Ig = grayscale image; ∇2 = Laplacian operator (second-order derivative for sharpness); Sx/Sy = Sobel horizontal/vertical edge responses; Mi,j = Sobel edge magnitude at (i,j) ; μM = mean edge magnitude; DCT(⋅) = Discrete Cosine Transform; Bk = k -th 8×8 image block; N=1000 (sampled blocks for efficiency); R = residual noise (image minus non-local means denoised version); pb = normalized frequency of Local Binary Pattern (LBP) bin b ; Cov(⋅,⋅) = covariance; Var(⋅) = variance; R/G/B = red/green/blue color channels; pv = normalized pixel value frequency for chrominance channels (Cr/Cb); Ei,j = Canny edge detector output (0=non-edge, 255=edge); Kurt(⋅) = kurtosis (tailedness of intensity distribution).
Figure 2: Stability ( Ss , left) and discriminability ( Sd , right) scores of image features for fake image detection. Based on the thresholds, set features to Core Feature(green), Usable Feature(orange), red Unstable High-Discrim(red), and Unusable Feature(grey).
Figure 3: Density distributions of core features from original images (Laplacian variance, Sobel magnitude mean/std over BigGAN/CRN/Glide (UniversalFakeDetect). Blue = real images, red = fake images.
Figure 4: Image-text pairs with original image captions and enhanced by core features’ description.
Method
Ref
Midjourney
SDv1.4
SDv1.5
ADM
GLIDE
Wukong
VQDM
BigGAN
mAcc
UniFD
CVPR 2023 Ojha and others (2023)
93.9
96.4
96.2
71.9
85.4
94.3
81.6
90.5
88.8
NPR
CVPR 2024 Tan and others (2024b)
81.0
98.2
97.9
76.9
89.8
96.9
84.1
84.2
88.6
FreqNet
AAAI 2024 Tan and others (2024a)
89.6
98.8
98.6
66.8
86.5
87.3
75.8
81.4
86.8
FatFormer
CVPR 2024 Liu and others (2024)
92.7
100.0
99.9
75.9
88.0
99.9
98.8
55.8
88.9
C2pClip (Trump.Biden)
AAAI 2025 Tan and others (2025)
82.2
95.1
95.5
95.1
98.9
98.7
93.8
98.3
94.7
C2pClip (Deepfake.Camera)
AAAI 2025 Tan and others (2025)
88.2
90.9
97.9
96.4
99.8
98.8
96.5
98.7
95.8
Table 2: Comparison with state-of-the-art methods on the GenImage dataset, where the best confidence threshold for each method is selected on an isolated 20% validation split by maximizing the F1-score and then fixed when reporting test accuracy. The highest accuracy in each column is highlighted in bold. mAcc denotes the average accuracy across all evaluated generative models.
Method
Ref
DALLE
Glide
ADM
LDM
mAcc
100_10
100_27
50_27
100_10
100_27
50_27
100
200
200_cfg
CNNDet Wang and others (2020a)
CVPR 2020
51.8
53.3
53.0
54.2
54.9
51.9
52.0
51.6
52.8
FreDect Frank and others (2020)
ICML 2020
57.0
53.6
50.4
52.0
53.4
56.6
56.4
56.5
54.5
F3Net Qian and others (2020)
ECCV 2020
71.6
88.3
87.0
88.5
69.2
74.1
73.4
80.7
79.1
LGrad [28]
CVPR 2023
88.5
89.4
87.4
90.7
86.6
94.8
94.2
95.9
90.9
UnivFD Ojha and others (2023)
CVPR 2023
89.5
90.1
90.7
91.1
75.7
90.5
90.2
77.3
86.9
Table 3: Comparison with state-of-the-art methods on the test subsets of the UniversalFakeDetect benchmark in terms of Accuracy (%). The evaluated settings cover four generation sources, including DALLE, Glide, ADM, and LDM, with different sampling configurations as specified in the column headers. The best result in each setting is highlighted in bold. PPM-CLIP Wang and others (2026) , a recent CVPR 2026 detection method, is included as a contemporary baseline. The mean accuracy (mAcc) is computed across all evaluated settings.
AI-generated image detection identifies synthetic or AI-edited images and develops methods that generalize across generators, domains, and image transformations.
The rapid advancement of AI generated content (AIGC) has blurred the boundaries between real and synthetic images, exposing the limitations of existing deepfake detectors that often overfit to specific generative models. This adaptability crisis calls for a fundamental reexamination of the intrinsic physical characteristics that distinguish natural from AI-generated images. In this paper, we address two critical research questions: (1) What physical features can stably and robustly discriminate AI generated images across diverse datasets and generative architectures? (2) Can these objective pixel-level features be integrated into multimodal models like CLIP to enhance detection performance while mitigating the unreliability of language-based information? To answer these questions, we conduct a comprehensive exploration of 15 physical features across more than 20 datasets generated by various GANs and diffusion models. We propose a novel feature selection algorithm that identifies five core physical features including Laplacian variance, Sobel statistics, and residual noise variance that exhibit consistent discriminative power across all tested datasets. These features are then converted into text encoded values and integrated with semantic captions to guide image text representation learning in CLIP. Extensive experiments demonstrate that our method achieves state-of-the-art performance on multiple Genimage benchmarks, with near-perfect accuracy (99.8%) on datasets such as Wukong and SDv1.4. By bridging pixel level authenticity with semantic understanding, this work pioneers the use of physically grounded features for trustworthy vision language modeling and opens new directions for mitigating hallucinations and textual inaccuracies in large multimodal models.
Mei Qiu, Jianqiang Zhao, Yanyun Qu
SDIC Intelligence Information Technology Co., Ltd. · Xia’men University, China
Vision-Language Large Models (VLLMs) are promising for AI-generated image (AIGI) detection because they can produce both a prediction and a natural-language output. However, most existing VLLM-based detectors primarily fine-tune the language side while giving limited attention to low-level visual forensic cues. They also often depend on manually crafted prompts or human-annotated rationales, which limits scalability.We present UC-VLM, a unified multi-stage framework for AIGI detection that relies solely on binary supervision. UC-VLM first identifies effective instruction variants automatically. It then reuses the same binary label within a multi-stage training framework: (i) a visual discrimination objective that strengthens sensitivity to non-semantic forensic cues, and (ii) a label-conditioned generation objective that uses the binary label to supervise textual outputs. This design turns weak binary supervision into a shared supervision signal for both the visual pathway and the language output. Our key novelty is a unified multi-stage binary-supervised framework that consistently reuses the same authenticity labels for visual adaptation and label-conditioned text generation, while leveraging automatically optimized instructions to reduce prompt sensitivity without requiring human-written rationales or hand-crafted prompts.Experiments show that UC-VLM achieves 96.1% average accuracy on GenImage, exceeding the strongest prior result by 4.6%, and obtains 69.6% / 77.9% accuracy on Chameleon under ProGAN / SDV1.4 training, surpassing the best baseline by 11.2% / 15.3%, respectively.
Generalization remains a critical bottleneck in AI-generated image detection. Because many modern generators are proprietary or adversarially modified, existing detectors overfit to the low-level textural patterns of accessible training data, resulting in severe failures on unseen domains. Conventional regularization techniques (e.g., L1/L2 norms, Dropout) apply indiscriminate parametric constraints and fail to provide the domain-invariant structure necessary for cross-generator robustness. To address this, we propose Feature-Augmented Implicit Regularization (FAIR). FAIR introduces an orthogonal, macro-structural prior, specifically, Scene Composition Structure (SCS), during training to geometrically constrain the model's optimization trajectory. By augmenting the primary feature space with domain-invariant SCS features, FAIR explicitly penalizes texture-biased shortcut learning. Crucially, this structural prior is entirely discarded at inference, yielding a smoothed, generalized decision boundary with zero architectural or computational overhead. Extensive evaluations across five massive benchmarks demonstrate that integrating FAIR into state-of-the-art detectors significantly improves cross-generator generalization, boosting accuracy by up to 8.04% and establishing new state-of-the-art robustness in zero-shot transfer scenarios.
Md Redwanul Haque, Manzur Murshed, Manoranjan Paul +1
Deakin University, Burwood, VIC, Australia · Charles Sturt University, Bathurst, NSW, Australia