acceptodds
Withdrawn from ICLR 2027

Local Binary Pattern is a Good Descriptor for AI-Generated Image Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Advancements in image generation technologies have raised significant concerns about their potential misuse, such as producing misinformation and deepfakes. Therefore, there is an urgent need for effective methods to detect AI-generated images. Recent studies reveal that texture is an important cue for fake image detection since the generator typically correlates the values of nearby pixels and cannot generate as strong texture contrast as real data. However, these methods compute the image texture only by simply summing the residual of four directions. In this paper, we find that Local Binary Pattern (LBP), which is a good texture descriptor, can capture intrinsic forgery clues. To this end, we proposed an LBP-based AI-generated image Detector, named as LBPDbased. To fully leverage the LBP descriptor of images, we also proposed an LBP-guided AI-generated image detector LBPDguided, which combines LBP descriptor with image features by the proposed global refinement module from both spatial and channel perspectives. Extensive experiments validate the effectiveness of the proposed method. It also achieves state-of-the-art results across various evaluation benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-generated Image Detection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jzYUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the detection of AI-generated images, proposing the use of traditional Local Binary Patterns (LBP) to extract texture features and constructing both an LBP-only detector and a dual-branch detector that fuses RGB features. The latter employs a Global Refinement Module (GRM) to facilitate feature interaction across spatial and channel dimensions. The study achieves strong cross-generator detection results on multiple benchmarks, yielding the key insight that simple LBP representations retain significant discriminative power. However, further clarification is needed regarding how this approach differs from existing LBP-based detection methods and whether this finding holds true for contemporary, powerful generative models.

Critical strengths

S1. The experiments are extensive, and the comparative analysis is comprehensive. The paper evaluates the method across multiple datasets, various T2I generation models, and GANs, while incorporating recent detection baselines; overall, the experimental work is robust. Ablation studies regarding different LBP variants, encoders, and fusion modules further facilitate an understanding of each component's contribution to performance.

S2. The performance of the simple LBP-only detector is of research interest. Compared to detection methods relying on complex feature extraction or fusion architectures, the ability to achieve strong cross-generator performance using only LBP feature maps is a result worthy of further investigation. Explaining why this simple approach works—and under what conditions it outperforms complex schemes—would further clarify the paper's contribution.

Critical weaknesses

W1. Limited methodological novelty; lack of explanation for the advantages of the simple approach. Existing works such as Multi-modal Texture Fusion Network and MSAFNet have already fused LBP with other features for synthetic image detection. This paper fails to directly compare against these highly relevant methods or explain why a simple LBP-only detector achieves such strong results. A comparison under a unified experimental setup is recommended to clarify whether the performance gains stem from LBP itself, the encoder, or the training configuration.

W2. Lack of validation against state-of-the-art generative models. While the experiments are extensive, they do not include representative generative models from 2026. The LBP discrepancies relied upon in this paper may reflect texture artifacts specific to older generators and might not generalize to current systems. It is recommended to supplement the study with LBP distribution analyses, LBP-only detection results, and fusion-based detection results for newer models to verify the continued practical relevance of the core findings.

W3. Qualitative evidence and figure presentation require improvement. Current visualizations illustrate LBP differences but do not clearly indicate which local patterns drive the detection. Including localized close-ups of matched image pairs, examples from state-of-the-art generative models, and failure cases is recommended. Additionally, the layout—featuring excessive text wrapping around figures—hinders readability; reorganizing the placement and sizing of figures is advised.

Questions for authors

Q1. Does LBP remain effective on 2026 state-of-the-art text-to-image models?
The method appears to rely on local texture artifacts, which stronger generators may have substantially reduced. Could the authors evaluate LBP distributions and LBP-only detection on representative 2026 models, keeping the detector fixed?

Q2. Why is a simple LBP-only detector so competitive?
Could the authors compare against prior methods that incorporate LBP into more complex pipelines, such as Multi-modal Texture Fusion Network and MSAFNet, under matched conditions, and explain whether—and why—the simpler approach performs better?

I would consider raising my score if these experiments demonstrate strong LBP-only generalization to current generators and clarify its advantages over prior complex approaches.

Reviewer_CGwuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates Local Binary Patterns (LBP) for AI-generated image detection. Through visualizations, histograms, and distribution-distance measurements, the authors examine differences in local texture patterns between real and generated images. They introduce LBPD_based, which uses LBP maps alone, and LBPD_guided, which combines RGB and LBP features through a Global Refinement Module (GRM) comprising bidirectional spatial attention and cross-branch channel gating. Experiments cover four detection benchmarks, with additional analyses of input representations, fusion components, encoders, image degradation, and inference time.

Critical strengths
  1. The authors identify and validate the usefulness of an established LBP encoding for AI-generated image detection. In particular, LBPD_based achieves competitive performance on multiple benchmarks using a fixed LBP transformation, a ResNet-34 encoder, and a classification head. This demonstrates that a straightforward texture representation and detector architecture can be effective for this task.
  2. The methodological progression is clear, moving from the definition and variants of LBP to distribution analyses, a single-branch detector, and a two-branch fusion model. The detector pipelines and the main computations involved in spatial attention and channel gating are supported by diagrams or equations, making it relatively easy to understand how the models use local texture information and combine it with RGB features.
  3. The evaluation covers multiple generators and four detection benchmarks. The full model achieves average accuracies of 98.7%, 95.1%, and 96.4% on GenImage, DRCT-2M, and Co-Spy-Bench, respectively. Ablations of input representations, LBP variants, and fusion components support the method’s effectiveness. Experiments with an alternative RGB encoder also suggest that the benefits of fusion are not limited to a single encoder configuration. The image-degradation experiments and inference-time measurements in the appendix further broaden the evaluation.
Critical weaknesses
  1. Some results are inconsistent with the text. If Table 1 averages the eight subsets equally, ESSP and Effort should score approximately 89.7% and 90.1%, rather than 90.6% and 91.1%. Table 2 also contradicts the claim that guided performs best on every subset except SDv1-DR. The claimed gain of at least four percentage points on Co-Spy-Bench should refer specifically to average accuracy over external baselines. Figure 4’s Diff definition and display scales, and the statistical calculations and sampling stability underlying Figure 5 and Table 4, also need clarification.
  2. The ablations do not fully distinguish the contributions of LBP representations, encoders, and fusion. RGB-only and LBP-only in Table 6 use different encoders, while the full GRM concatenates four feature vectors without clearly reporting the dimensions and parameter counts of each configuration. RGB/LBP comparisons using the same encoder and fusion baselines of comparable capacity would help determine whether the gains come from the input representation, fusion design, or model capacity.
Questions for authors
  1. Please clarify the averaging in Table 1 and verify the ESSP and Effort averages, the claims associated with Tables 2–3, and the DRCT-2M category count in the appendix. Please consistently mark the best, second-best, and tied results.
  2. Does FFT Diff in Figure 4 show a difference between magnitude spectra or the FFT magnitude of a difference map? Please clarify the color scales, frequency bands, energy calculations, and legends.
  3. Please clarify Figure 5’s vertical axis, aggregation across images and channels, and sampling stability, along with the JSD and EMD calculations in Table 4. Does “more consistent patterns” in Figure 1 mean repetitive textures rather than consistency between raw images and LBP maps or across RGB channels?
  4. Please report feature dimensions and parameter counts for each fusion configuration, and include fusion baselines of comparable capacity and RGB/LBP comparisons using the same encoder. Is the LBP encoder fully fine-tuned, and what LoRA settings are used for RGB?
  5. Why should retaining orientation information and grouping nonuniform patterns in the 59-category LBP encoding benefit generated-image detection?
  6. Does LBPD_based match the LBP-only configuration and training procedure in Table 6? Please distinguish the gains from RGB and GRM, clarify the timing conditions and inclusion of LBP computation, and compare efficiency under identical conditions.
  7. Please report key results across multiple random seeds and explain why guided trails IAPL on SDv1-DR and achieves relatively low absolute accuracy on SDv2-DR and SDXL-DR.
Reviewer_9LxLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates Local Binary Patterns (LBP) for AI-generated image detection. The authors argue that real and generated images differ in local microtexture and that LBP exposes these differences as useful classification features. They propose two detectors: LBPD_based classifies LBP maps using ResNet-34, while LBPD_guided combines LBP features with DINOv2 RGB features through a Global Refinement Module (GRM) containing spatial cross-attention and channel gating. The paper reports strong results on four benchmarks and supports the approach with distribution analyses, comparisons of input representations, and component ablations.

Critical strengths
  1. The simple baseline produces a potentially valuable empirical result.LBPD_based achieves average accuracies of 95.2%, 93.0%, and 93.4% on GenImage, DRCT-2M, and Co-Spy-Bench, respectively.
  2. The method exploits texture cues directly from the input image without requiring diffusion reconstruction.Both detectors compute LBP from the input image and perform classification without extracting errors through diffusion reconstruction. This reduces dependence on a reconstruction model and provides a straightforward approach to AI-generated image detection.
  3. The evaluation covers multiple generators and model categories. Experiments include diffusion models and GANs, image degradations, and an alternative RGB encoder. This coverage is more convincing than evaluation on a single generator.
Critical weaknesses
  1. The potential influence of image-processing differences is insufficiently controlled. The paper attributes the effectiveness of LBP to microtexture differences between real and AI-generated images. However, compression and resampling can also alter the local pixel relationships encoded by LBP. The manuscript does not sufficiently clarify whether these processing conditions are matched between real and generated images, making it difficult to determine how much of the detection performance reflects generation-related cues rather than possible processing differences. The authors should clarify the preprocessing settings and provide controlled comparisons under matched processing conditions to better support their interpretation.
  2. The sources and experimental settings of the baseline results are insufficiently specified. The paper provides basic training settings for the proposed method and specifies the training data and evaluation protocols in Tables 1–3. However, it remains unclear whether the baseline results are taken from prior publications or reproduced by the authors. The relevant preprocessing, training, and model-selection settings are also insufficiently documented. This makes it difficult to assess how much of the performance difference is attributable to the methods themselves. The authors should clarify the sources and key experimental conditions of the baseline results, including setting differences that may affect the comparisons.
  3. Limited fusion baselines leave the advantage of GRM’s specific design insufficiently established.Tables 6 and 7 support the effectiveness of SR, CR, and the complete GRM in the tested settings. However, comparisons with alternative fusion schemes of similar parameter counts, such as concatenation followed by an MLP, are missing. Such comparisons under consistent training conditions, together with parameter counts and computational costs for the fusion modules, would help distinguish the benefits of GRM’s specific design from those of added model capacity.
Questions for authors
  1. Preprocessing and potential confounding factors. Could the authors clarify whether real and generated images are matched in terms of resolution, resampling methods, encoding formats, and compression settings during training and evaluation, and whether LBP is computed before or after these preprocessing operations? What evidence shows that the advantage of LBP persists under matched processing conditions, beyond the image-degradation tests in Appendix A.3?
  2. Contribution of the GRM design. Tables 6 and 7 demonstrate improvements over direct feature concatenation. Does GRM retain its advantage over a fusion baseline with a comparable parameter count, such as concatenation followed by an MLP, using the same encoders and training settings? A controlled comparison, together with the parameter counts and computational costs of the fusion modules, would help determine whether the gains arise from the proposed spatial/channel interactions or from additional model capacity.
  3. Contribution relative to closely related work. Prior work has explored combining LBP with RGB and other texture representations for AI-generated image detection. Could the authors clarify the specific contributions of the present work beyond these existing approaches, particularly in terms of LBP analysis, feature representation, and fusion design? Which findings or experimental results support these additional contributions?

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

Related papers

Discussion (0)

Sign in to comment.