acceptodds
Withdrawn from ICLR 2027

MiLD: Mining Informative Low-Bit Cues for Training-Free AI-Generated Image Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Low-bit planes have been shown to contain discriminative information for AI-generated image detection. We further observe that the Hamming states induced by low-bit morphological residuals provide useful real/fake cues. However, the Hamming-state occupancy is a static image-level statistic and is sensitive to JPEG compression and Gaussian noise. This motivates us to use Hamming states as spatial cues rather than as the final detection score, and to construct a stable, robust, and efficient training-free detector. We propose MiLD, which selects channel-disagreement locations from low-bit morphological residuals and builds a masked low-bit reversal. MiLD then compares the frozen encoder features of the original and reversed images, using their feature discrepancy as the detection score. Across five benchmarks, MiLD improves the mean AUROC over RIGID by 2.78% while achieving the highest efficiency score. Extensive ablations further demonstrate the stability, transferability, and robustness of MiLD.

Primary area:
generative models
Keywords:
Training-Free, AI-Generated Image Detection

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_UYdXOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MiLD, a training-free detector for AI-generated images based on low-bit morphological residuals. The method identifies channel-disagreement locations using Hamming states, applies a masked low-bit reversal, and uses the feature discrepancy between the original and reversed images from a frozen encoder as the detection score. However, the experimental comparison is not sufficiently up to date for an ICLR 2027 submission. Although the paper cites recent works from 2026, the main experiments do not provide a comprehensive comparison with the latest detectors. Moreover, MiLD does not achieve the best overall performance: its five-benchmark mean AUROC is 80.14%, which is 4.41 points lower than WaRPAD. In addition, the technical novelty appears incremental, and the theoretical analysis does not fully explain why the proposed perturbation should distinguish real and generated images.

Critical strengths

1.The paper presents a simple and efficient training-free detector. MiLD only requires two forward passes of a frozen encoder and does not train an additional detector on target-domain real/fake images.

2.The use of low-bit morphological residuals and Hamming states to locate channel-disagreement locations is interesting and provides a clear motivation for the masked low-bit reversal.

Critical weaknesses

1.The experiments do not sufficiently compare against the latest relevant methods. Most of the main baselines are from 2024–2025. Although the paper discusses more recent works such as RDD and RAID, these methods are not included in the main performance comparison.

2.The relationship to existing low-bit and bit-reversal methods is not sufficiently clarified. The paper discusses LOTA, LIDA, and RAID, but the fundamental differences between these methods and MiLD are not fully established.

3 The generalization results are mixed and do not fully support the broad claims.

4The theoretical analysis provides limited insight into the actual detection mechanism. The Taylor expansion shows that the feature discrepancy measures the sensitivity of the frozen representation along the masked low-bit reversal direction. However, it does not explain why generated images should consistently produce a larger response than real images.

Questions for authors

1.Why do the authors not compare their method with more recent baselines? 2.Recent works, such as Effort (ICML25) HSIC(ICLR26), adopt accuracy (ACC) as the evaluation metric. Relying solely on AUC is insufficient to fully reflect detector performance.

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

This paper studies training-free AI-generated image detection. It observes that low-bit planes contain generation traces, and more specifically that the Hamming states of low-bit morphological residuals differ systematically between real and generated images. Based on this, the authors propose MiLD, which applies signed bidirectional morphology to the low-bit image, selects channel-disagreement locations, complements the low bits only at those locations while preserving the high bits, and scores an image by the feature discrepancy between a frozen DINOv2 encoder's representations of the original and reversed images. Across five benchmarks, MiLD reaches 80.14% mean AUROC, and achieves the highest efficiency score at 1.73s per 100 images. However, its performance remains below the strongest baseline WaRPAD.

Critical strengths

1.Exploring low-bit generation traces and making them usable via morphology.

Critical weaknesses
  1. The method is largely empirical, and the morphology/residual design is under-justified. The signed bidirectional operator, the 3×33\times3 kernel, b=2b=2, and the Hamming-state set 1,2{1,2} are fixed from the initial observation rather than derived from a principled argument; the theoretical section only interprets the final score as a local Jacobian sensitivity and explicitly states that it does not explain why the morphology-selected direction separates real from generated images. This makes the pipeline look post-hoc and leaves its core mechanism unvalidated, which directly weakens soundness.

  2. Robustness is weak and the claimed performance-efficiency trade-off is not convincing. At JPEG quality 50, MiLD drops to 80.18 AUROC, the lowest among the non-reconstruction baselines. At Gaussian noise σ=32\sigma=32, it drops to 77.66, below RIGID, MINDER, WePe and ConV, and only above the collapsed WaRPAD. Its five-benchmark mean is 80.14, below WaRPAD's 84.55, and it is not the best on any of the five benchmarks. Since the paper's central claim is a robust and efficient training-free detector, these results make the robustness claim hard to sustain and leave the method's practical value unclear.

Questions for authors
  1. Design rationale and mechanism validation. The pipeline fixes signed bidirectional opening/closing, a 3×33\times3 kernel, but the paper justifies these choices mainly by the initial observation. Could the authors state the principled reason for each.

  2. Robustness and operating regime. At JPEG quality 50, MiLD is the weakest non-reconstruction method, and under (\sigma=32) noise it falls below RIGID, MINDER, WePe and ConV; its mean AUROC is also below WaRPAD on all five benchmarks. Given this sensitivity, could the authors add an experiment that randomly perturbs 1–2 pixels per image and reports the resulting change in AUROC? This would directly quantify the method's sensitivity to small perturbations and clarify whether the low-bit cue is stable enough for practical use.

  3. Presentation. Some presentation should be optimized. For example, Figure 1 shows the outputs of the different morphological tools, but not whether these outputs separate real from generated images; the reader cannot tell which tool provides the discriminative signal that motivates the pipeline. Could the authors mark the key difference between real and generated images.

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

This paper proposes MiLD, a training-free AI-generated image detector based on morphology guided low bit reversal. The authors observe that intermediate Hamming states of low bit morphological residuals, corresponding to RGB channel disagreement patterns, exhibit systematic differences between real and generated images. Since the occupancy statistics of these patterns are sensitive to JPEG compression and noise, they are used as spatial cues rather than direct detection scores. Specifically, MiLD selectively reverses the low order bits at the identified locations and uses the feature discrepancy between the original and modified images, measured by a frozen pretrained encoder, as the detection score. Across multiple benchmarks, MiLD consistently outperforms RIGID at a comparable inference cost, highlighting its ability to balance detection performance and computational efficiency while maintaining competitive results across different pretrained backbones.

Critical strengths
  1. The paper presents a novel empirical observation that RGB channel disagreement patterns in low-bit morphological residuals differ between real and AI-generated images, highlighting their potential as discriminative cues for AI-generated image detection.
  2. The paper proposes a novel training-free detection method that applies selective perturbations guided by low-bit morphology and uses the resulting feature discrepancy from a frozen pretrained encoder as the detection score.
  3. The paper conducts extensive comparisons with existing training-free methods across multiple benchmarks and validates the effectiveness of its key components through ablation studies. It also provides a comprehensive evaluation of computational efficiency, compatibility with different pretrained backbones, and robustness to image corruptions.
Critical weaknesses
  1. The paper does not sufficiently justify adopting a training-free approach despite its potential performance gap with training-based detectors. Without comparisons to training-based methods or clear application scenarios, it remains unclear whether eliminating detector training outweighs the loss in detection performance.
  2. Although the paper distinguishes MiLD from generic reconstruction and perturbation methods, it does not clearly explain their limitations or why targeted perturbations based on low bit cues are necessary. Its methodological novelty relative to prior methods exploiting frequency cues also remains unclear.
  3. While MiLD offers a favorable balance between accuracy and computational cost, its detection performance remains below WaRPAD. In particular, its near-random AUROC of 51.77% on DailyBench raises concerns about generalization to recent generative models and practical applicability.
  4. Although ablation studies support morphology guided mask selection, improvements over raw low bit based masking are limited. The paper does not sufficiently explain what unique discriminative information morphological channel disagreement provides or the underlying mechanism distinguishing real from AI-generated images.
Questions for authors
  1. What practical benefits or application scenarios justify adopting a training-free approach despite its potential performance gap with training-based detectors? Under what conditions would eliminating detector training outweigh the loss in detection performance?
  2. What specific limitations of generic reconstruction or perturbation methods does MiLD address? How does its targeted use of low-bit cues differ from existing methods such as WaRPAD, which also exploit frequency domain information?
  3. Why does MiLD achieve only 51.77% AUROC on DailyBench? Could the authors analyze whether certain generators or manipulation methods contribute more to this degradation and discuss how generalization to recent generative models could be improved?
  4. Given the limited improvement over raw low-bit masking, what unique discriminative information does morphological channel disagreement capture? Could the authors provide theoretical or empirical evidence explaining why these patterns distinguish real from AI-generated images and their advantages over alternative cues?

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

Related papers

Discussion (0)

Sign in to comment.