acceptodds
Withdrawn from ICLR 2027

: Simplifying Data Distributions for Effective AI-Generated Image Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The rapid progress of image generation technology has heightened concerns about misuse, including misinformation and deepfakes, creating an urgent need for reliable AI-generated image (AIGI) detection. Although state-of-the-art detectors based on inversion and reconstruction achieve strong performance in the ideal setting, where the inversion and reconstruction models are the same, their performance drops significantly when these two models differ. All previous works on diffusion-based AIGI detection overlook an important prior, that the noise of generated images from various models is drawn from a standard Gaussian distribution . We propose to exploit this prior in order to improve the generalization and robustness of AIGI detection. We first perform a series of analyses on the distributional dynamics of the diffusion inversion process. Our analysis shows that the prediction errors of the denoising model accumulate at each inversion step, preventing the real image distribution converging to in the noise space. To further deepen our analysis, we employ a Chi-squared test and a Bayes classifier to identify the distributions to which real and generated image patches converge. Our results show that the distributions of generated and real image patches converge to two very simple forms: the former approximates , while the latter becomes a mixture of a non-standard Gaussian distribution and a distribution with small . Motivated by these findings, we propose Inversion Noise Classification (INC), a straightforward classification method that operates directly on . Extensive experiments across diverse diffusion models and settings show that INC consistently outperforms state-of-the-art AIGI detectors, with notable superiority in robustness to input distortions and generalizability across inversion methods.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AIGI detection, Gaussian noise prior, Diffusion inversion

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_wYFiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Reconstruction-based AI-generated image detectors invert an input image and use reconstruction errors for detection, but their performance can degrade when the generation and inversion models differ or when images undergo distortions. This paper instead distinguishes real and generated images directly in the terminal noise representation zTz_T obtained through diffusion inversion. The authors argue that generated-image representations approach a standard Gaussian distribution, whereas real-image representations exhibit different statistical properties due to inversion mismatch. Their proposed Inversion Noise Classification (INC) method encodes an image with a VAE, obtains zTz_T through inversion, and classifies it using ResNet-18, without image reconstruction or reconstruction-error computation. The approach is supported by an error-accumulation analysis, patch-energy statistics, and QDA classification experiments. Experiments across multiple models and evaluation settings demonstrate strong detection performance.

Critical strengths
  1. A simple but powerful idea. The paper offers an appealing perspective: classify the terminal inversion representation zTz_T rather than the initial latent representation z0z_0. Beyond proposing this design, the authors visualize and analyze its distributional properties, providing compelling empirical support for the choice of representation.

  2. An effective and efficient approach. Despite its simplicity, INC achieves competitive detection performance against strong baselines. By avoiding reconstruction and operating directly on a compact latent representation, it also offers computational advantages over methods such as FakeInversion.

  3. Useful theoretical motivation. I find the analysis in Section 4.1 helpful in explaining why terminal inversion representations can be informative for detection. In particular, it describes how additional real-image mismatch terms contribute to the terminal latent gap while the contribution of the initial gap diminishes. Although this does not directly guarantee improved classification, it provides a useful rationale for the proposed approach.

Critical weaknesses
  1. Unclear training protocol and scope of generalization. It is unclear whether a separate classifier is trained for every T2I–I2I combination in Tables 2 and 3, or whether a classifier trained on one combination is evaluated on others without retraining. Appendix J suggests that the training and test sets share the same T2I–I2I provenance, but the retraining and weight-sharing protocol is not explicit. Please clarify this setup for both INC and the trainable baselines. If classifiers are retrained for each combination, these results demonstrate effectiveness across model combinations rather than generalization of a fixed detector to unseen generators or inversion models.

  2. Missing FakeInversion comparison in Table 4. Table 4 evaluates robustness by training on clean images and testing on distorted images, but compares only DIRE and INC. FakeInversion achieves the highest mean accuracy in Table 3, yet is omitted from Table 4 without explanation. Its results under the same conditions are necessary to assess robustness relative to this strong baseline. Table 5 does not resolve this omission because it evaluates the opposite direction of distribution shift.

  3. Checkpoint selection using test accuracy. Appendix J states that the checkpoint with the highest test accuracy is selected for each training run. Applying the same protocol to all trainable baselines may provide some consistency, but it does not ensure an unbiased comparison: the test set is used for model selection, and the resulting selection benefit may differ across methods. Please clarify whether “test” actually refers to a separate validation set. Otherwise, checkpoints should be selected using validation data and evaluated on an independent test set.

  4. The accuracy–efficiency trade-off needs stronger evidence. INC performs worse than FakeInversion in some settings, including the mean accuracy in Table 3. This gap may be acceptable given INC's computational advantages. While Appendix K.1 provides an initial cost comparison, the paper should make this trade-off more explicit and support it with additional measurements, particularly end-to-end inference latency including feature extraction, throughput, and memory usage under matched hardware and evaluation conditions.

Questions for authors

See Weaknesses Section

Reviewer_9wdAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Inversion Noise Classification (INC) for AI generated image detection. It VAE-encodes an image, applies deterministic DDIM inversion to obtain the terminal latent z_T, and trains a ResNet-18 to classify real versus generated directly on that latent, without reconstruction or pixel-space features. The experiments cover nine T2I/I2I diffusion models, and the authors claim that INC consistently outperforms existing SOTA detectors.

Critical strengths
  1. The paper derives a linear recursion for the noise-prediction error of deterministic DDIM inversion, showing that the initial real-versus-generated latent gap is contracted at each step while the terminal gap is dominated by the accumulated real-image-specific mismatch terms.

  2. Building on this recursion, patch-wise chi-squared energy tests, predicted-noise mean/covariance statistics and QDA show that the z_T of generated images approaches a standard Gaussian, whereas the z_T of real images forms a mixture of a non-standard Gaussian component and a near-zero-energy atom.

Critical weaknesses
  1. The primary concern is that the abstract and conclusion claim INC "consistently outperforms state-of-the-art" detectors, whereas many reported results do not support this claim. FakeInversion outperforms the proposed INC in most comparisons in Table 3 and in most columns of Table 21. These results do not substantiate the authors' claim.

  2. The presentation of the experimental results is inconsistent and difficult to parse. In several tables the semantics of boldface and underlining are left undefined. Table 5 is particularly confusing, containing multiple bold entries within a single row and multiple bold and underlined entries within a single column.

  3. The novelty and the incremental contribution of the method are limited. Reduced to its components, INC is FakeInversion with the reconstruction stage removed and the classifier relocated from pixel-space/reconstruction features to the terminal inversion latent z_T; FakeInversion (Cazenavette et al., CVPR 2024) already performs detection using Stable Diffusion inversion features, and in the paper's own Table 2 FakeInversion is stronger in most clean settings.

  4. Several state-of-the-art reconstruction-error-based detectors are cited but never compared, including DRCT (Chen et al., ICML 2024) and LARE² (Luo et al., CVPR 2024), all of which operate under closely related reconstruction protocols.

  5. Benchmarking by the per-row maximum (MAX) over inversion models is an unfair protocol: a method that attains 100% on a single easy configuration receives MAX = 100, which renders the comparison uninformative. Moreover, only balanced accuracy is reported, without TPR at a fixed false-positive rate or the false-positive rate on pristine photographs, which are the standard operating points for forensic deployment.

  6. The manuscript does not define the terminology used in its tables and contains numerous typographical errors; the authors should revise it carefully.

Questions for authors
  1. Have the authors tested closed-source high-quality generators such as GPT-Image-2 and Nano-Banana-2.0?

  2. Have the authors evaluated real-image sets beyond LAION-10k and ImageNet-10k, and how do metrics such as FPR and AUC compare across different real sets? The paper reports only balanced accuracy, while the false-positive rate on pristine photographs and the full ROC are exactly the metrics that determine whether a forensic detector is deployable.

Reviewer_fsZMOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Inversion Noise Classification (INC), which directly classifies the terminal inversion latent $z_T$ for AI-generated image detection and motivates it through a distributional analysis showing that zTz_T provides a simpler and more separable representation than z0z_0. While the method is simple and empirically strong, its novelty is somewhat incremental and the current evaluation leaves concerns about benchmark coverage, comparison fairness, and potential dataset-specific shortcuts.

Critical strengths

1. Clear and insightful motivation: The paper provides a simple and intuitive perspective on AI-generated image detection by showing that fully inverted latent representations$z_T$ are more prior-aligned and statistically easier to separate than the original latent space z0z_0.

2. Easy-to-follow presentation: The paper is generally well organized and easy to follow. The motivation, theoretical analysis, statistical observations, and final detection method are presented in a coherent progression, making the overall methodology and main message easy to understand.

Critical weaknesses

Weaknesses

1. Limited novelty: FakeInversion and DTADhave already explored AI-generated image detection using diffusion inversion and inversion/noise representations. The main novelty of this work lies in directly classifying zTz_Tand providing a distributional interpretation, while the algorithmic contribution itself appears relatively incremental.

2. Concerns about effectiveness and comparison fairness: The experiments show that directly classifying the original latent representation z0z_0 already achieves strong detection performance on the authors' datasets, while the improvement from z0z_0to zTz_T is relatively modest. Moreover, INC trains a supervised ResNet classifier, making direct accuracy comparisons with training-free methods such as DTAD and AEROBLADE not entirely fair.

3. Insufficient experimental validation: The evaluation is conducted mainly on the authors' self-constructed LAION-10k and ImageNet-10k datasets, without experiments on widely used public benchmarks such as GenImage. This makes it difficult to rule out the influence of dataset construction and source-specific biases.

4. Potential test-set leakage in the evaluation protocol: Appendix J states that the data are split only into training and test sets with an 80/20 ratio, and that the checkpoint with the highest test accuracy is selected. This effectively uses the test set for model selection and may lead to overly optimistic reported performance.

Questions for authors
  1. How does the proposed method perform on commonly used benchmarks in this field, such as GenImage and Chameleon?

  2. How does the method perform on the SynthBuster[1] dataset? The real images in SynthBuster (RAISE-1k) are uncompressed. To the best of my knowledge, DIRE has been shown to rely heavily on JPEG-related shortcuts in real images. Since DIRE achieves strong performance across nearly all experiments in this paper, I am concerned that the proposed method may also be exploiting JPEG/compression artifacts in the real images rather than genuinely learning generation-specific cues.

[1] Synthbuster: Towards Detection of Diffusion Model Generated Images. Bammey, et al. OJSP 2024.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.