Pith. sign in

REVIEW 4 major objections 6 minor 65 references

Reversing an image's bit-plane order amplifies the faint artifacts left by AI generators, and a millisecond detector built on that representation beats existing methods on more than 40 benchmarks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 16:14 UTC pith:VYZSR5P7

load-bearing objection Empirically clean and fast detector buried under an invalid theory section and an abstract that contradicts its own Table 2. the 4 major comments →

arxiv 2607.28974 v1 pith:VYZSR5P7 submitted 2026-07-31 cs.CV cs.AI

RAID: Towards Robust AI-Generated Image Detection with Bit-Reversed Images

Renxi Cheng , Jie Gui , Hongsong Wang This is my paper
classification cs.CV cs.AI
keywords AI-generated image detectiondeepfake detectionbit-plane decompositionbit-reversed imageforensic artifact amplificationgradient-based patch selectioncross-generator generalizationzero-shot detection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that AI-generated images can be spotted by the statistics of their low-order bit-planes. The authors introduce the bit-reversed image, a reversible rearrangement in which the eight bit-planes of each channel are recombined with reversed weights so that noise and fine detail dominate; in fake images this brings out forensic artifacts that are invisible in the original. They then select the single highest-gradient 32×32 patch and classify it with a lightly modified ResNet-50, obtaining a pipeline called RAID. The paper reports top accuracy on over 40 benchmarks, across cross-generator, cross-dataset, and zero-shot settings, with 23.5 million parameters and millisecond inference. If those results hold, a compact, fast, and general detector does not need large reconstruction or pretrained diffusion modules.

Core claim

RAID's central claim is that reversing the order of the eight bit-planes of an image—so that the least significant planes receive the largest reconstruction weights—amplifies the forensic micro-structures generative models leave behind while suppressing the semantic content that normally masks them. The paper argues that real images carry physical sensor noise in the low bit-planes, whereas generated images show structured or unnatural randomness there because they lack such noise; the bit-reversed image turns that difference into a visible artifact. From this representation, gradient-based patch selection finds the most artifact-rich patch, and an adapted ResNet-50 (with reduced early strid

What carries the argument

The bit-reversed image (BRI) is the central object: an image recomposed from its eight bit-planes using the weight vector [2^7, 2^6, ..., 2^0] instead of [2^0, ..., 2^7], so low-order planes dominate the resulting pixel values. Because this reordering is reversible and parameter-free, it acts as a fixed preprocessing transform meant to suppress semantic content and amplify fine artifacts. Two additional mechanisms carry the pipeline: a gradient-based patch-selection score (the L1 norm of convolutions with horizontal, vertical, and two diagonal kernels, maximized over patches) that isolates the most artifact-laden 32×32 patch, and an ImageNet-pretrained ResNet-50 adapted by reducing early str

Load-bearing premise

Everything rests on the empirical premise that generated images have detectable irregularities in their low-order bit-planes—unlike the sensor noise of real photos—and that reversing the bit-plane order exposes those irregularities rather than destroying them.

What would settle it

Take a set of real and AI-generated test images, replace the lowest three bit-planes (bits 0–2) of each image with uniform random noise or zero them out, then run RAID's trained classifier. If accuracy stays high, the detector is not actually using low-bit-plane artifacts; if accuracy collapses toward chance, the mechanism the paper claims is confirmed.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A 23.5M-parameter network that processes a single 32×32 patch of a bit-reversed image suffices for state-of-the-art detection, implying real-time and low-resource deployment without huge pretrained backbones.
  • Because the representation is parameter-free and the classifier transfers across generators (e.g., trained on Stable Diffusion V1.4 and tested on ADM, VQDM, BigGAN, and Midjourney), one RAID model can address generators it never saw during training.
  • Zero-shot performance of 86.3% average accuracy on GenImage, built only from ImageNet real images, suggests that detectors can be prepared for future generators without needing examples of their output.
  • Ablation results on GenImage (98.4% vs. 87.7% without BRI and 74.8% without patch selection) indicate both the bit-reversed representation and the gradient-based patch selection are each responsible for a large part of the gain.
  • Degradation experiments under Gaussian blur and JPEG compression show the BRI signal survives moderate degradation better than comparable single-patch and frequency-based methods, extending the method's practical reach to compressed social-media imagery.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: the proof's additive decomposition (Eq. 6) ignores binary carries, so the bit-planes of a sum are not simply the sum of bit-planes; the empirical detector could be right even where the theoretical justification is approximate, and the representation itself, not the proof, is the more durable contribution.
  • Inference: because the detector's signal is concentrated in the low-order bit-planes, a generator that explicitly whitens or regularizes those planes' statistics is the natural adversary to test next; the paper does not consider such an attack.
  • Inference: the same reversible bit-reversal transform could be transferred to other tasks where low-bit statistics matter—steganalysis, camera-model attribution, compression-history identification—though the paper only evaluates image and video forgery detection.
  • Inference: the paper closes by stating its own limitation—the classifier is a standard ResNet, adapted only by stride changes—so the headroom in a purpose-built architecture, and how it interacts with the bit-reversed representation, is left untested.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes RAID, a pipeline for AI-generated image detection that constructs a bit-reversed image from the input, selects a single informative patch via a gradient-based score, and classifies that patch with an adapted ResNet-50. The authors also introduce two new benchmarks (GID for generated images, GVD for images extracted from generated videos) and report extensive cross-generator, cross-dataset, zero-shot, robustness, and efficiency experiments. The central empirical claim is that RAID achieves state-of-the-art accuracy on over 40 benchmarks while being about 100 times faster than previous detectors and using only 23.5M parameters.

Significance. If the empirical results are taken at face value, the paper's contribution is significant: a compact, fast detector that generalizes across unseen generators and datasets would be practically valuable, and the bit-reversal representation is an interesting and simple idea. The strengths are the breadth of the experimental evaluation, the ablations isolating the BRI and GBPS components (Tables 7, 10, 11), the robustness study (Table 8), the efficiency comparison (Table 9), and the public code release. The reported cross-generator and cross-dataset numbers are striking. However, the paper overstates its own results in the abstract and in §4.2, and the claimed theoretical justification in Section 3.3 is mathematically invalid as written. These issues are fixable, but they affect how the contribution is framed and must be corrected before publication.

major comments (4)
  1. [§1, §4.2, Table 2] The abstract and §4.2 claim that RAID 'surpasses all prevailing methods' on AIGCDB and 'outperforms existing methods on over 40 benchmarks.' This is contradicted by the paper's own Table 2: on AIGCDB, the average accuracy of C2P is 96.2% while RAID is 93.5%. AIGCDB is one of the two primary benchmarks, so the flagship SOTA claim is internally inconsistent. Please qualify the claim to the benchmarks/subsets where RAID actually leads, correct the AIGCDB sentence, and provide a verifiable count of the 'over 40 benchmarks.' As written, the central contribution is overstated by the paper's own numbers.
  2. [§3.3, Eqs. (6)–(8), Propositions 1–3] The theoretical analysis is mathematically invalid. Eq. (6) writes the bit-planes of I as sums of bit-planes of S, A, and R, and Eq. (7) assumes that bit-reversal commutes with addition. But bit-plane extraction is nonlinear because of carries in binary arithmetic: the bit-planes of a sum are not the sum of the individual bit-planes. Consequently, Propositions 1–3 do not follow from the stated assumptions. The algorithm itself never relies on the S/A/R decomposition, so the empirical results may stand, but the claimed mathematical proof of validity must be removed, corrected, or explicitly reframed as informal motivation.
  3. [§4.1, §4.2, Table 6] The zero-shot experimental protocol is underspecified. The text says that only real ImageNet images are used, features are obtained after global average pooling, and distances are evaluated, but it does not state the distance metric, the decision threshold, the number of reference real images, how patch selection is applied at test time, or how the comparison baselines RIGID, AEROBLADE, and Manifold are treated under the same protocol. Please provide a reproducible zero-shot protocol; otherwise the zero-shot results in Table 6 cannot be independently verified.
  4. [§3.1, Related Work] The novelty claim is internally inconsistent. The Introduction states that marrying bit-planes with AI-generated image detection 'has not been studied yet,' while Related Work cites LOTA [47] as a bit-plane-guided AI-generated image detection method and says the proposed approach is 'similar to LOTA.' Please clarify the relationship to LOTA and revise the novelty statement accordingly.
minor comments (6)
  1. [Table 2] The table header is ambiguous: the first numeric column is AIGCDB average, while the remaining columns are GenImage subsets, but the header places 'AIGCDB GenImage' side by side. Please label the columns explicitly so readers do not mistake the AIGCDB average for a GenImage subset.
  2. [§4.2, Table 3] The text says the model trained on GLIDE attains an averaged performance of 87.0%, but Table 3 lists the GLIDE row average as 78.7%. Please correct the text or the table.
  3. [Eq. (2)] There is a typographical corruption in the displayed formula for the gradient score: the norm expression contains stray ' ' characters. Please fix the typesetting.
  4. [§5] The conclusion contains a grammatical error: 'this paper, we studies' should be 'this paper, we study.'
  5. [Tables 10 and 11] The circled numbers (1⃝, 2⃝, etc.) are used without a clear definition in the captions. Please state in each caption that they denote different bit-order/weight variants and refer to Section 4.3.
  6. [Abstract and §2] The name 'WANX 2.1' is inconsistently spaced ('W ANX 2.1' in Table 1). Please standardize.

Circularity Check

3 steps flagged

Empirical RAID results are not circular, but the paper's three validity propositions reduce to their own assumptions; theoretical justification is self-definitional.

specific steps
  1. self definitional [Sec. 3.3, Proposition 1, Eq. (6)-(8)]
    "According to the principle of image bit-plane decomposition, the I, S and R of a bit-forward image can all be decomposed as follows: I(x, y) = Σ 2^k · s_k^c + Σ 2^k · a_k^c + Σ 2^k · r_k^c, where s_k^c, a_k^c and r_k^c represent the k-th bit-plane in the R channel of S, A and R."

    This equation assumes bit-plane decomposition is linear under addition: the bit-planes of I are treated as the bit-planes of S plus A plus R. In binary arithmetic this is false because of carries. More importantly, the proof then defines 'semantic content' as Σ2^k s_k and observes that reversing the weights 2^k → 2^{7−k} makes Eq. (8) hold. That inequality is a statement about the chosen weights, i.e. about the definition of bit-reversal, not a property of images. The conclusion that semantic content is weakened is imported by the additive model rather than derived.

  2. self definitional [Sec. 3.3, Proposition 2, Eq. (9)]
    "During the generation of a fake image, due to the influence of processes such as upsampling and denoising, we have: I(x, y) = S(x, y) + δ, where δ ≠ 0. Since R(x, y) = 0 for generated images, it follows that A(x, y) ≠ 0 ... Thus, Ã(x, y) ≠ 0, which means the micro-structure of fake images is inevitably present."

    In model (5), A is defined as the non-semantic, non-noise residual. The proof assumes a fake image has a nonzero residual δ and sets R=0 for generated images, so A=δ follows by definition. The claim that fake images inevitably contain forensic microstructure is therefore the assumption I=S+δ rewritten in the notation of (5), not an independent result. It does not establish that bit-reversal exposes anything, since no actual bit-plane computation is used in the argument.

  3. self definitional [Sec. 3.3, Proposition 3, Eqs. (10)-(11)]
    "Since ˜S(x, y) is very small after bit-reversion and ˜R(x, y) = 0, the gradients ∇˜S and ∇˜R are negligible. Therefore, ∇˜I is primarily dominated by ∇˜A ... Σ˜A = S · Σ∇˜I ... It is evident that the sum of forensic micro-structure in certain region achieve the maximum value if and only if the gradient of the patch ∇˜I reaches its maximum."

    The proof makes the quantity to be maximized (ΣÃ) proportional to the gradient score (Σ∇Ĩ) by assuming ∇Ĩ ≈ ∇Ã and then writing ΣÃ = S·Σ∇Ĩ. GBPS is defined in Eqs. (2)-(4) as selecting the patch with the largest gradient score. Thus the claimed equivalence 'maximum artifacts iff maximum gradient' is built into the assumption that the artifact field's gradient equals the image gradient; the 'if and only if' is a restatement of that equality, not a proof that high-gradient patches in bit-reversed images carry forensic artifacts.

full rationale

RAID's empirical contribution is not circular: Tables 2-6 evaluate the pipeline on unseen generators and datasets, and ablations (Table 7) show the BRI and GBPS modules matter; those numbers are not forced by the theoretical section. However, the paper's 'theoretical analysis from the mathematical perspective' is partially circular. Proposition 2 assumes a fake image has a nonzero residual and R=0, then concludes A≠0, which is a restatement of model (5). Proposition 3 assumes ∇Ĩ ≈ ∇Ã and then writes ΣÃ = S·Σ∇Ĩ, making the artifact sum proportional to the gradient score by construction, so the claimed 'if and only if' is an assumption. Proposition 1's Eq. (6) treats bit-plane decomposition as linear under addition, which is false due to carries; the inequality (8) then merely restates that the chosen reversed weights are oppositely ordered. These are self-definitional problems in the validity argument, though the detector's performance against external benchmarks does not depend on them. The self-citation to LOTA [47] is used only as confirmation after in-paper figures and ablations, so it is not load-bearing. The paper's own limitation statement (Sec. 5, standard ResNet) is not a circularity issue. The apparent contradiction between the abstract's 'over 40 benchmarks' claim and Table 2's AIGCDB result (RAID 93.5 vs C2P 96.2) is an accuracy/claim issue, not a circularity issue. Score 5 reflects partial circularity in the theoretical justification while the central empirical claim remains independently tested.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The method's empirical pipeline is simple and reproducible, but the paper's claimed theoretical support rests on an additive decomposition and a linearity assumption for bit-planes that are not valid; the choice of weights and patch size are hand-set, and the learnable-weight variant performs worse. The new datasets GID/GVD are introduced but not released, so the benchmark claims cannot be independently checked.

free parameters (5)
  • bit-plane weight vector w = [128, 64, 32, 16, 8, 4, 2, 1]
    Hand-chosen weights for bit-reversed image construction (Section 3.1); Table 11 searches alternatives, and variant 5 reaches 99.3% avg while the reported default is 98.4%.
  • patch size = 32×32
    Chosen for the adapted ResNet-50; no justification given in Section 3.2.
  • number of selected patches = 1
    GBPS selects a single patch with max gradient (Eq. 4), a design choice not derived.
  • input resize resolution = 256×256
    Implementation detail in Section 4.1; affects which artifacts survive.
  • learnable weights variant (Table 11 8) = trained (avg 82.2%)
    A learnable weighting was tested and underperforms, so hand-set weights are used.
axioms (5)
  • ad hoc to paper Image decomposes additively as I = S + A + R with A≈0 for real and R=0 for generated images
    Eq. (5), Section 3.3; this decomposition is assumed to justify bit-plane analysis but is not derived or independently measured.
  • ad hoc to paper Bit-plane decomposition is linear under addition of S, A, R
    Eq. (6) assumes the bit-planes of the sum equal the sum of bit-planes; false due to binary carries. This invalidates Propositions 1 and 2.
  • ad hoc to paper Sum of forensic microstructure in a region is proportional to the integral of its gradient (Prop 3)
    Eq. (11) is dimensionally inconsistent; the claim that ∑ A = S·∑∇A is not justified.
  • domain assumption Artifacts from generators concentrate in low-order bit-planes
    Section 3.1, supported by visual evidence (Fig. 1, 3) and LOTA [47], but not by a quantitative model.
  • domain assumption ImageNet-pretrained ResNet-50 features transfer to bit-reversed images
    The adapted ResNet-50 is initialized from ImageNet weights (Section 4.1); this transfer is assumed to work.
invented entities (2)
  • Forensic microstructure A(x,y) no independent evidence
    purpose: Named component of the additive image model used to attribute artifacts in generated images
    Defined to be ≈0 for real and ≠0 for fake (Section 3.3); no independent measurement; Prop 2's 'proof' is tautological.
  • Bit-reversed image (BRI) independent evidence
    purpose: A representation that amplifies low-bit-plane artifacts by reversing bit-plane weights
    It is a deterministic function of the image, reproducible by anyone from the description; evidence for its utility is the empirical accuracy gain in Table 7 (w/o BRI drops to 87.7%).

pith-pipeline@v1.3.0-daily-deepseek · 17905 in / 14041 out tokens · 104786 ms · 2026-08-03T16:14:04.862221+00:00 · methodology

0 comments
read the original abstract

The rapid advancement of image generation models has made it increasingly difficult for people to distinguish AI-generated images from real ones. To prevent the potential risks associated with the misuse of fake images, AI-generated image detection has gained significant attention. Existing methods neglect the inherent differences between real and fake images, thus lacking robustness and generalization ability. In this work, we innovatively investigate AI-generated image detection using bit-planes, and introduce the bit-reversed image. We propose a simple yet effective pipeline consisting of construction of bit-reversed images, gradient-based patch selection and a convolutional classifier. Besides, we provide a theoretical analysis from the mathematical perspective to demonstrate the validity of our approach. We also introduce two challenging datasets for AI-generated image detection. Extensive experiments verify the effectiveness of our approach across different settings, including cross-generator generalization, cross-dataset generalization and zero-shot performance. Without bells and whistles, our approach outperforms existing methods on over 40 benchmarks, and is nearly 100 times faster than counterparts. The code is at https://github.com/renxi-seu/RAID.

Figures

Figures reproduced from arXiv: 2607.28974 by Hongsong Wang, Jie Gui, Renxi Cheng.

Figure 1
Figure 1. Figure 1: Comparison of bit-reversed images between real and fake images. Noticeable artifacts appear in certain regions of reversed images in fake images, while the noise in bit-reversed images of real images tends to be more naturally distributed. domain between generated and real images. These artifacts manifest as unnatural directional properties in the spectrum of synthetic images, with diffusion-generated imag… view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline for the proposed method. Given an RGB image, our approach first extracts the bit-reversed image, then heuristically selects a patch, and finally performs fake classification using the adapted ResNet-50 (↓2 indicates a 2× downsampling of spatial resolution in feature learning). As for patch-based methods, PatchCraft [64] regards the difference between the patches with the highest and low￾est divers… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of sample distributions between real and fake images of different bit-planes. For low bit-planes, especially bit-0, bit-1, and bit-2, the distributions of real and fake images are clearly separable, and different types of generators also exhibit distinct sample distributions. In contrast, for high bit-planes, the distributions of real and fake images heavily overlap. order of eight bit-planes be… view at source ↗
Figure 4
Figure 4. Figure 4: Visualizations of bit-planes, bit-reversed and bit-forward images for real and fake RGB images. k = {0, 1, . . . , 7} is index of bit-planes. For bit-forward or bit-reversed images, k denotes the positions of the left circular shift operation. For AI-generated images, different values of k result in varying artifacts. defines the set of gradient directions. The convolution ker￾nels gx, gy, gxy and gyx are … view at source ↗
Figure 5
Figure 5. Figure 5: Evaluation of generalization capability. Four competitive methods and our approach are trained on eight subsets (corresponding to eight subplots) and evaluated on eight subsets (corresponding to eight dimensions of each subplot) of GenImage [65]. stream GAN and Diffusion based generators, includ￾ing BigGAN (BigG)[2], Midjourney (Midj)[34], Wukong (Wuk)[53], Stable Diffusion V1.4 (SDV4)[41], Stable Dif￾fusi… view at source ↗
Figure 6
Figure 6. Figure 6: Visualizations of AI-generated images and the corresponding probabilities of small patches predicted as fake. The images are resized to 256×256, with a patch size of 32×32. The predicted fake probabilities are at the center of each patch on the right of RGB images, with more intense blue colors representing patches that are successfully predicted as fake. tains only 23.5 million parameters and runs in mill… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

65 extracted references · 7 linked inside Pith

  1. [1]

    Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 8

    2023

  2. [2]

    Large scale gan training for high fi- delity natural image synthesis

    Andrew Brock et al. Large scale gan training for high fi- delity natural image synthesis. InInternational Conference on Learning Representations, 2018. 7

    2018

  3. [3]

    Manifold in- duced biases for zero-shot and few-shot detection of gener- ated images

    Jonathan Brokman, Amit Giloni, Omer Hofman, Roman Vainshtein, Hisashi Kojima, and Guy Gilboa. Manifold in- duced biases for zero-shot and few-shot detection of gener- ated images. InInternational Conference on Learning Rep- resentations, 2025. 9

    2025

  4. [4]

    Hiding data in im- ages by simple lsb substitution.Pattern recognition, 37(3): 469–474, 2004

    Chi-Kwong Chan and Lee-Ming Cheng. Hiding data in im- ages by simple lsb substitution.Pattern recognition, 37(3): 469–474, 2004. 1, 3

    2004

  5. [5]

    Videocrafter2: Overcoming data limitations for high-quality video diffusion models.arXiv preprint arXiv:2401.09047, 2024

    Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models.arXiv preprint arXiv:2401.09047, 2024. 8

    Pith/arXiv arXiv 2024

  6. [6]

    A single simple patch is all you need for ai-generated image detection.arXiv preprint arXiv:2402.01123, 2024

    Jiaxuan Chen, Jieteng Yao, and Li Niu. A single simple patch is all you need for ai-generated image detection.arXiv preprint arXiv:2402.01123, 2024. 1, 3, 5, 7, 8, 9, 10

    Pith/arXiv arXiv 2024

  7. [7]

    Ed4: Explicit data- level debiasing for deepfake detection.IEEE Transactions on Image Processing, 34:4618–4630, 2025

    Jikang Cheng, Ying Zhang, Qin Zou, Zhiyuan Yan, Chao Liang, Zhongyuan Wang, and Chen Li. Ed4: Explicit data- level debiasing for deepfake detection.IEEE Transactions on Image Processing, 34:4618–4630, 2025. 1

    2025

  8. [8]

    Intriguing properties of syn- thetic images: from generative adversarial networks to dif- fusion models

    Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of syn- thetic images: from generative adversarial networks to dif- fusion models. InIEEE / CVF Computer Vision and Pattern Recognition Conference, pages 973–982, 2023. 2

    2023

  9. [9]

    On the detection of synthetic images generated by diffusion mod- els

    Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Gio- vanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion mod- els. InIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2023. 2

    2023

  10. [10]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255. Ieee, 2009. 6, 8

    2009

  11. [11]

    Diffusion models beat gans on image synthesis.NeurIPS, 2021

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis.NeurIPS, 2021. 7

    2021

  12. [12]

    Fourier spectrum discrepancies in deep network generated images

    Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images. InAdvances in Neural Information Processing Systems, pages 3022–3032, 2020. 1

    2020

  13. [13]

    An image steganography approach based on k- least significant bits (k-lsb)

    Omar Elharrouss, Noor Almaadeed, and Somaya Al- Maadeed. An image steganography approach based on k- least significant bits (k-lsb). InIEEE international confer- ence on informatics, IoT, and enabling technologies (ICIoT), pages 131–135. IEEE, 2020. 1

    2020

  14. [14]

    A chaotic image encryption algorithm based on 3-d bit-plane permutation.Neural Computing and Applications, 31(11): 7111–7130, 2019

    Zhi-hua Gan, Xiu-li Chai, Dao-jun Han, and Yi-ran Chen. A chaotic image encryption algorithm based on 3-d bit-plane permutation.Neural Computing and Applications, 31(11): 7111–7130, 2019. 1, 3

    2019

  15. [15]

    Toward generalizable forgery detection and reasoning.IEEE Transactions on Image Processing, 35:3395–3410, 2026

    Yueying Gao, Dongliang Chang, Bingyao Yu, Haotian Qin, Muxi Diao, Lei Chen, Kongming Liang, and Zhanyu Ma. Toward generalizable forgery detection and reasoning.IEEE Transactions on Image Processing, 35:3395–3410, 2026. 2

    2026

  16. [16]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. InConference on Neural Information Processing Systems, 2014. 1

    2014

  17. [17]

    Vec- tor quantized diffusion model for text-to-image synthesis

    Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vec- tor quantized diffusion model for text-to-image synthesis. In IEEE / CVF Computer Vision and Pattern Recognition Con- ference, pages 10696–10706, 2022. 8

    2022

  18. [18]

    Towards lossless im- plicit neural representation via bit plane decomposition

    Woo Kyoung Han, Byeonghun Lee, Hyunmin Cho, Sunghoon Im, and Kyong Hwan Jin. Towards lossless im- plicit neural representation via bit plane decomposition. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 2269–2278, 2025. 1, 3

    2025

  19. [19]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016. 5, 7, 9

    2016

  20. [20]

    Rigid: A training-free and model-agnostic framework for ro- bust ai-generated image detection.arXiv preprint arXiv:2405.20112, 2024

    Zhiyuan He, Pin-Yu Chen, and Tsung-Yi Ho. Rigid: A training-free and model-agnostic framework for ro- bust ai-generated image detection.arXiv preprint arXiv:2405.20112, 2024. 9

    Pith/arXiv arXiv 2024

  21. [21]

    Cogvideo: Large-scale pretraining for text-to-video generation via transformers.arXiv preprint arXiv:2205.15868, 2022

    Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers.arXiv preprint arXiv:2205.15868, 2022. 8

    Pith/arXiv arXiv 2022

  22. [22]

    Secret lies in color: Enhancing ai-generated images detection with color distribution anal- ysis

    Zexi Jia, Chuanwei Huang, Yeshuang Zhu, Hongyan Fei, Xi- aoyue Duan, Zhiqiang Yuan, Ying Deng, Jiapei Zhang, Jin- chao Zhang, and Jie Zhou. Secret lies in color: Enhancing ai-generated images detection with color distribution anal- ysis. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 13445–13454, 2025. 2

    2025

  23. [23]

    Glff: Global and local feature fusion for ai-synthesized im- age detection.IEEE Transactions on Multimedia, 26:4073– 4085, 2023

    Yan Ju, Shan Jia, Jialing Cai, Haiying Guan, and Siwei Lyu. Glff: Global and local feature fusion for ai-synthesized im- age detection.IEEE Transactions on Multimedia, 26:4073– 4085, 2023. 2

    2023

  24. [24]

    Countering malicious deepfakes: Survey, battleground, and horizon.International Journal of Com- puter Vision, 2022

    Felix Juefei-Xu, Run Wang, Yihao Huang, Qing Guo, Lei Ma, and Yang Liu. Countering malicious deepfakes: Survey, battleground, and horizon.International Journal of Com- puter Vision, 2022. 1

    2022

  25. [25]

    Any-resolution ai- generated image detection by spectral learning

    Dimitrios Karageorgiou, Symeon Papadopoulos, Ioannis Kompatsiaris, and Efstratios Gavves. Any-resolution ai- generated image detection by spectral learning. InProceed- ings of the Computer Vision and Pattern Recognition Con- ference, pages 18706–18717, 2025. 2

    2025

  26. [26]

    Text2video-zero: Text- to-image diffusion models are zero-shot video generators

    Levon Khachatryan, Andranik Movsisyan, Vahram Tade- vosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text- to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954–15964, 2023. 8

    2023

  27. [27]

    Dig- ital image stabilizing algorithms based on bit-plane match- ing.IEEE Transactions on Consumer Electronics, 44(3): 617–622, 1998

    Sung-Jea Ko, Sung-Hee Lee, and Kyung-Hoon Lee. Dig- ital image stabilizing algorithms based on bit-plane match- ing.IEEE Transactions on Consumer Electronics, 44(3): 617–622, 1998. 1, 3

    1998

  28. [28]

    Bit-plane based re- versible data hiding in encrypted images using multi-level blocking with quad-tree.IEEE Transactions on Multimedia, 26:4722–4735, 2023

    Rajeev Kumar, Ajay K Sharma, et al. Bit-plane based re- versible data hiding in encrypted images using multi-level blocking with quad-tree.IEEE Transactions on Multimedia, 26:4722–4735, 2023. 3

    2023

  29. [29]

    Detecting generated images by real images

    Bo Liu, Fan Yang, Xiuli Bi, Bin Xiao, Weisheng Li, and Xinbo Gao. Detecting generated images by real images. In European Conference on Computer Vision, pages 95–110. Springer, 2022. 7, 9

    2022

  30. [30]

    Forgery-aware adaptive transformer for generalizable synthetic image detection

    Huan Liu, Zichang Tan, Chuangchuang Tan, Yunchao Wei, Jingdong Wang, and Yao Zhao. Forgery-aware adaptive transformer for generalizable synthetic image detection. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 10770–10780, 2024. 7

    2024

  31. [31]

    Global tex- ture enhancement for fake face detection in the wild

    Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global tex- ture enhancement for fake face detection in the wild. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8060–8069, 2020. 7, 9

    2020

  32. [32]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In International Conference on Computer Vision, 2021. 7

    2021

  33. [33]

    Lareˆ 2: Latent reconstruction error based method for diffusion-generated image detection

    Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding. Lareˆ 2: Latent reconstruction error based method for diffusion-generated image detection. InIEEE / CVF Com- puter Vision and Pattern Recognition Conference, pages 17006–17015, 2024. 2, 7, 8, 9, 10

    2024

  34. [34]

    Midjourney.https://www.midjourney

    Midjourney. Midjourney.https://www.midjourney. com/home/, 2022. 7

    2022

  35. [35]

    Genvidbench: A challenging benchmark for detecting ai-generated video, 2025

    Zhenliang Ni, Qiangyu Yan, Mouxiao Huang, Tianning Yuan, Yehui Tang, Hailin Hu, Xinghao Chen, and Yunhe Wang. Genvidbench: A challenging benchmark for detecting ai-generated video, 2025. 8

    2025

  36. [36]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741, 2021

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741, 2021. 7

    Pith/arXiv arXiv 2021

  37. [37]

    Towards uni- versal fake image detectors that generalize across genera- tive models

    Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards uni- versal fake image detectors that generalize across genera- tive models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480– 24489, 2023. 7, 9, 10

    2023

  38. [38]

    A little bit more: Bitplane-wise bit-depth recovery.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9718– 9724, 2021

    Abhijith Punnappurath and Michael S Brown. A little bit more: Bitplane-wise bit-depth recovery.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9718– 9724, 2021. 3

    2021

  39. [39]

    Thinking in frequency: Face forgery detection by mining frequency-aware clues

    Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. InEuropean Conference on Computer Vision, pages 86–103. Springer, 2020. 7

    2020

  40. [40]

    Aer- oblade: Training-free detection of latent diffusion images using autoencoder reconstruction error

    Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aer- oblade: Training-free detection of latent diffusion images using autoencoder reconstruction error. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9130–9140, 2024. 9

    2024

  41. [41]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. InIEEE / CVF Computer Vision and Pattern Recognition Conference, 2022. 1, 7

    2022

  42. [42]

    Learning on gradients: Generalized arti- facts representation for gan-generated images detection

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized arti- facts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 12105–12114, 2023. 1, 7, 9

    2023

  43. [43]

    Frequency-aware deepfake de- tection: Improving generalizability through frequency space domain learning

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Frequency-aware deepfake de- tection: Improving generalizability through frequency space domain learning. InProceedings of the AAAI Conference on Artificial Intelligence, pages 5052–5060, 2024. 1, 2

    2024

  44. [44]

    Rethinking the up-sampling op- erations in cnn-based generative network for generalizable deepfake detection

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling op- erations in cnn-based generative network for generalizable deepfake detection. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 28130–28139, 2024. 7

    2024

  45. [45]

    C2p-clip: Inject- ing category common prompt in clip to enhance generaliza- tion in deepfake detection

    Chuangchuang Tan, Renshuai Tao, Huan Liu, Guanghua Gu, Baoyuan Wu, Yao Zhao, and Yunchao Wei. C2p-clip: Inject- ing category common prompt in clip to enhance generaliza- tion in deepfake detection. InProceedings of the AAAI Con- ference on Artificial Intelligence, pages 7184–7192, 2025. 7

    2025

  46. [46]

    Training data-efficient image transformers & distillation through at- tention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv ´e J´egou. Training data-efficient image transformers & distillation through at- tention. InInternational Conference on Machine Learning,

  47. [47]

    LOTA: Bit-planes guided ai-generated image detection

    Hongsong Wang, Renxi Cheng, Yang Zhang, Chaolei Han, and Jie Gui. LOTA: Bit-planes guided ai-generated image detection. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 17246–17255, 2025. 3

    2025

  48. [48]

    Unsupervised domain adaptation-based cross-type deepfake image detection.IEEE Transactions on Image Processing, 35:4411–4424, 2026

    Qin Wang, Xiaofeng Wang, Zinian Liu, Ningning Bai, Minghua Zhao, and Shanmin Pang. Unsupervised domain adaptation-based cross-type deepfake image detection.IEEE Transactions on Image Processing, 35:4411–4424, 2026. 1

    2026

  49. [49]

    Cnn-generated images are surprisingly easy to spot

    Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8695–8704, 2020. 7, 9

    2020

  50. [50]

    Videofactory: Swap at- tention in spatiotemporal diffusions for text-to-video gener- ation.arXiv preprint arXiv:2305.10874, 2023

    Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap at- tention in spatiotemporal diffusions for text-to-video gener- ation.arXiv preprint arXiv:2305.10874, 2023. 8

    Pith/arXiv arXiv 2023

  51. [51]

    Dire for diffusion-generated image detection

    Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. InIEEE / CVF Computer Vision and Pattern Recognition Conference, pages 22445–22455, 2023. 1, 2, 7, 9, 10

    2023

  52. [52]

    Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation

    Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623–7633, 2023. 8

    2023

  53. [53]

    Wukong.https://xihe.mindspore.cn/ modelzoo/wukong, 2022

    Wukong. Wukong.https://xihe.mindspore.cn/ modelzoo/wukong, 2022. 7

    2022

  54. [54]

    Musev: Infinite-length and high fidelity virtual human video genera- tion with visual conditioned parallel denoising.arxiv, 2024

    Zhiqiang Xia, Zhaokang Chen, Bin Wu, Chao Li, Kwok-Wai Hung, Chao Zhan, Yingjie He, and Wenjiang Zhou. Musev: Infinite-length and high fidelity virtual human video genera- tion with visual conditioned parallel denoising.arxiv, 2024. 8

    2024

  55. [55]

    Are high-quality ai-generated images more dif- ficult for models to detect? InInternational Conference on Machine Learning, 2025

    Yao Xiao, Binbin Yang, Weiyan Chen, Jiahao Chen, Zijie Cao, ZiYi Dong, Xiangyang Ji, Liang Lin, Wei Ke, and Pengxu Wei. Are high-quality ai-generated images more dif- ficult for models to detect? InInternational Conference on Machine Learning, 2025. 3

    2025

  56. [56]

    A sanity check for ai- generated image detection

    Shilin Yan, Ouxiang Li, Jiayin Cai, Yanbin Hao, Xiaolong Jiang, Yao Hu, and Weidi Xie. A sanity check for ai- generated image detection. InInternational Conference on Learning Representations, 2025. 7, 8, 9

    2025

  57. [57]

    All patches matter, more patches better: En- hance ai-generated image detection via panoptic patch learn- ing.arXiv preprint arXiv:2504.01396, 2025

    Zheng Yang, Ruoxin Chen, Zhiyuan Yan, Ke-Yue Zhang, Xinghe Fu, Shuang Wu, Xiujun Shu, Taiping Yao, Shouhong Ding, and Xi Li. All patches matter, more patches better: En- hance ai-generated image detection via panoptic patch learn- ing.arXiv preprint arXiv:2504.01396, 2025. 1, 3

    arXiv 2025

  58. [58]

    Mora: Enabling gen- eralist video generation via a multi-agent framework, 2024

    Zhengqing Yuan, Ruoxi Chen, Zhaoxu Li, Haolong Jia, Li- fang He, Chi Wang, and Lichao Sun. Mora: Enabling gen- eralist video generation via a multi-agent framework, 2024. 8

    2024

  59. [59]

    Towards universal ai-generated image detec- tion by variational information bottleneck network

    Haifeng Zhang, Qinghui He, Xiuli Bi, Weisheng Li, Bo Liu, and Bin Xiao. Towards universal ai-generated image detec- tion by variational information bottleneck network. InPro- ceedings of the Computer Vision and Pattern Recognition Conference, pages 23828–23837, 2025. 7

    2025

  60. [60]

    Detect- ing and simulating artifacts in gan fake images

    Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detect- ing and simulating artifacts in gan fake images. InIEEE in- ternational workshop on information forensics and security (WIFS), pages 1–6. IEEE, 2019. 1, 7

    2019

  61. [61]

    Reversible data hid- ing in encrypted images with asymmetric coding and bit- plane block compression.IEEE Transactions on Multimedia, 26:10174–10188, 2024

    Xianquan Zhang, Feiyi He, Chunqiang Yu, Xinpeng Zhang, Ching-Nung Yang, and Zhenjun Tang. Reversible data hid- ing in encrypted images with asymmetric coding and bit- plane block compression.IEEE Transactions on Multimedia, 26:10174–10188, 2024. 3

    2024

  62. [62]

    Learned lossless image compression based on bit plane slic- ing

    Zhe Zhang, Huairui Wang, Zhenzhong Chen, and Shan Liu. Learned lossless image compression based on bit plane slic- ing. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 27579–27588,

  63. [63]

    Breaking semantic arti- facts for generalized ai-generated image detection.Advances in Neural Information Processing Systems, 37:59570–59596,

    Chende Zheng, Chenhao Lin, Zhengyu Zhao, Hang Wang, Xu Guo, Shuai Liu, and Chao Shen. Breaking semantic arti- facts for generalized ai-generated image detection.Advances in Neural Information Processing Systems, 37:59570–59596,

  64. [64]

    Patchcraft: Exploring texture patch for efficient ai-generated image detection.arXiv preprint arXiv:2311.12397, 2023

    Nan Zhong, Yiran Xu, Sheng Li, Zhenxing Qian, and Xinpeng Zhang. Patchcraft: Exploring texture patch for efficient ai-generated image detection.arXiv preprint arXiv:2311.12397, 2023. 1, 2, 3, 5, 6, 7

    Pith/arXiv arXiv 2023

  65. [65]

    Genimage: A million-scale benchmark for de- tecting ai-generated image.Advances in Neural Information Processing Systems, 36, 2024

    Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for de- tecting ai-generated image.Advances in Neural Information Processing Systems, 36, 2024. 6, 7, 8, 9

    2024