acceptodds
Withdrawn from ICLR 2027

Spectral Tail Auxiliary Learning for AI-Generated Image Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

As generative image models evolve rapidly, the perceptual gap between generated and real images continues to narrow, making AI-generated image detection increasingly challenging. Many existing methods exploit frequency-domain cues for detection, typically described as frequency-domain artifacts or high-frequency discrepancies. However, the specific and recurring spectral regularities remain insufficiently understood and characterized. In this paper, we systematically analyze the one-dimensional radial log-power spectra of real and generated images. We find that generated images do not necessarily exhibit higher or lower energy across the entire spectrum or high-band range. Instead, their spectra deviate from the power-law decay and show an anomalous uplift in the ultra-high-frequency tail. We term this phenomenon spectral tail uplift. We further explain this phenomenon through nonlinear harmonic accumulation in trained generative models, suggesting that it can serve as a structural cue across generative architectures. Based on this observation, we propose Spectral Tail Auxiliary Learning (STAL), a frequency-domain auxiliary supervision framework for generalizable AI-generated image detection. STAL transfers spectral-tail cues from a tail-aware frequency teacher to a spatial detector during training, while all frequency-domain modules are discarded at inference time. Consequently, STAL introduces no inference overhead. Extensive experiments on 9 public datasets show that STAL achieves strong generalization and stability across generators, data distributions, and real-world scenarios.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-generated image detection, Image forensics, Generalizable detection, Frequency-domain analysis, Auxiliary learning

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

This paper investigates a recurring frequency-domain characteristic of AI-generated images, termed spectral tail uplift, where generated images exhibit anomalous energy increases in the ultra-high-frequency spectrum. The authors provide a theoretical explanation based on nonlinear harmonic accumulation and propose Spectral Tail Auxiliary Learning (STAL), which transfers frequency-domain information to a spatial detector through training-time auxiliary supervision. Experiments across nine benchmarks demonstrate strong generalization and robustness without additional inference overhead.

Critical strengths
  1. The phenomenon is characterized systematically across many generators and autoencoders, with consistent controlled experiments.

  2. Interesting spectral observation with theoretical motivation. The paper identifies spectral tail uplift as a recurring characteristic across different generative architectures. The theoretical analysis and controlled experiments provide a plausible explanation based on nonlinear harmonic accumulation. This offers a potentially valuable perspective on transferable forensic cues for AI-generated image detection.

Critical weaknesses
  1. Unclear relationship between spectral-tail supervision and inference-time robustness. The detector maintains high accuracy under blur and downsampling, which substantially suppress the high-frequency components motivating the proposed method. Although spectral-tail supervision may encourage more robust spatial representations, the paper provides no direct evidence for this mechanism. The improvements could instead arise from generic auxiliary supervision rather than the specific spectral-tail cue. Targeted frequency-band interventions and comparisons with non-tail frequency teachers would help clarify this distinction.

  2. Incomplete theoretical justification and insufficient differentiation from prior work. The theoretical results show that nonlinear activations can generate high-frequency harmonics, but do not establish why these necessarily produce a spectral uplift near the Nyquist limit. The analysis also does not adequately address aliasing in discretely sampled images, which provides a plausible alternative explanation. Furthermore, the relationship to prior studies of high-frequency artifacts and nonlinear aliasing is insufficiently discussed, weakening the paper’s novelty claims.

Questions for authors
  1. If the performance gains are attributed to spectral-tail supervision, why does the detector remain robust when the corresponding high-frequency information is suppressed? Does the spatial encoder retain sensitivity to the spectral tail, or does the auxiliary supervision instead encourage the learning of other, more robust features? Please provide targeted frequency-band interventions and feature-level analyses to clarify this mechanism.

  2. How much of the gain remains when the teacher is replaced by a random target, or by a spectral teacher of the same dimension built from a non-tail band? This would separate the contribution of the tail cue from that of generic auxiliary regularization.

  3. Since STAL distills frequency features into the spatial encoder, how does it compare with other training-time frequency supervision or distillation approaches, rather than only with detectors that use frequency inputs at inference?

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

The paper studies a localized uplift in the high-frequency tail of the radial power spectrum of generated images. The authors investigate its origin through harmonic analysis and decoder interventions, and propose Spectral Tail Auxiliary Learning (STAL) to incorporate this cue into a spatial detector during training. The frequency branch is discarded at inference. Experiments on nine benchmarks show improved cross-dataset detection, with additional studies on image degradations and individual model components.

Critical strengths
  1. The paper makes an interesting observation about a localized spectral pattern across different generators. Studying the shape of the spectral tail, rather than simply its energy, provides a useful perspective on frequency-based detection.
  2. The auxiliary supervision is simple and practically appealing. It improves the average balanced accuracy from 94.2% to 97.0% over the matched-backbone DDA* baseline while retaining only the spatial detector at inference. The improvement on ForenSynths is particularly substantial.
  3. The experimental coverage is reasonably broad. The nine benchmarks, component ablations, and degradation experiments provide complementary evidence for the method. The tail-only result also supports the usefulness of the proposed frequency supervision.
Critical weaknesses
  1. The analysis explains how harmonics can be generated, but not why they should produce the particular tail shape observed in the experiments. In particular, the filter gains in Equation 6 may suppress these components. The claim in lines 265 to 267 that they become the dominant tail contribution therefore needs further justification.
  2. The activation and weight interventions are difficult to interpret causally. Replacing all activations or randomizing the weights changes the decoder output substantially, including its reconstruction quality and overall spectrum. These experiments are consistent with the proposed explanation, but do not isolate it from other consequences of disrupting the decoder.
  3. I am unsure how to interpret several results in Table 3. A balanced accuracy of 0.2% for NPR on GoT would imply almost completely reversed predictions for both classes. Please verify the metric and score conventions. The criterion and validation data used for best-checkpoint selection in Appendix C.3 should also be stated.
  4. Table 2 shows that the tail head is useful, but does not establish whether the gain depends on the proposed uplift morphology. A comparable teacher using ordinary high-frequency statistics would make this distinction clearer. The predefined operators in Section 3.3 also need explicit definitions.
  5. Figure 2 gives aggregate uplift values, leaving the variation across individual images unclear. It would be useful to see the distributions for real and generated images, particularly across different real-image sources. This would help assess how consistently the proposed cue separates the two classes.
  6. The reported standard deviations describe variation across test subsets, rather than uncertainty across training runs. They should not be interpreted as evidence that small differences are reproducible. Repeated runs of the matched-backbone comparison would help establish the stability of the reported improvement.
Questions for authors
  1. What conditions are needed for the harmonics in Theorems 1 and 2 to produce a local increase in radial spectral power? Is the current analysis intended as a possible explanation or as an established account of the observed uplift?
  2. Could you provide the real-class and generated-class accuracies underlying NPR’s 0.2% result on GoT, together with the score direction and threshold? How were the checkpoints for STAL and DDA* selected?
  3. What operations constitute the tail head? Have you compared it with a similarly sized teacher using generic high-frequency statistics? This would help clarify whether the improvement specifically supports the proposed spectral observation.
Reviewer_eLiWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Spectral Tail Auxiliary Learning (STAL), a frequency-domain auxiliary supervision framework for AI-generated image detection. The authors identify Spectral Tail Uplift, a recurring increase in spectral power at the extreme high-frequency tail of generated images, and explain it through nonlinear harmonic generation and accumulation in trained generative models. STAL employs a Frequency Teacher combining a Tail Head and a Frequency Context Encoder, transferring spectral cues to a spatial detector through representation alignment. During inference, only the spatial detector is retained, eliminating additional frequency-domain computation. Experiments on nine public benchmarks demonstrate superior performance over existing methods, including one of the strongest detectors DDA with the same backbone, along with strong cross-generator generalization and robustness to post-processing.

Critical strengths
  1. The authors identify Spectral Tail Uplift, a novel frequency domain characteristic distinguishing real and AI-generated images. They systematically investigate this phenomenon across different generative architectures and provide theoretical analysis and controlled experiments to explain its underlying mechanism through nonlinear harmonic generation and accumulation.

  2. The proposed STAL framework introduces a novel auxiliary learning approach that transfers spectral information from a Frequency Teacher to a Spatial Detector through representation alignment. Notably, the frequency branch is removed during inference, improving detection performance without additional inference overhead.

  3. Extensive experiments across diverse generative architectures, including GANs, diffusion models, and autoregressive models, as well as in-the-wild datasets, demonstrate strong generalization and superior performance over competitive baselines, including DDA. The results also demonstrate robustness against common image post-processing operations.

  4. Ablation studies examine different frequency feature integration strategies, the contributions of key components, and performance across different backbones, providing empirical support for the effectiveness of the proposed framework.

Critical weaknesses
  1. Although Spectral Tail Uplift is systematically investigated across various generative models, the paper does not sufficiently clarify the sources and diversity of real images used in the spectral analysis. Given that MSCOCO is used for training, it remains unclear whether the observed spectral differences are influenced by characteristics specific to a particular real image dataset. Further analysis of spectral power characteristics and Spectral Tail Uplift distributions across diverse real image datasets would help establish the generality of the proposed finding.
  2. The proposed framework transfers frequency domain features learned by a Frequency Teacher to a Spatial Detector through representation alignment, but the effectiveness of this approach is not sufficiently explained. Although various methods exist for integrating spatial and frequency features, the specific feature fusion strategy used in the Direct Spatial-Frequency Inference baseline in Table 2 remains unclear. Furthermore, the absence of an ablation study excluding the alignment loss makes it difficult to isolate and assess the independent contribution of representation alignment.
  3. The extreme high-frequency region is defined using a fixed frequency range to capture Spectral Tail Uplift, but the rationale behind this choice is not clearly justified. In particular, it remains unclear whether the frequency range in which Spectral Tail Uplift occurs is consistent across different generative models. A clearer justification for the definition of the extreme high frequency region and the selection of its frequency boundary is therefore needed.
Questions for authors
  1. The paper analyzes Spectral Tail Uplift across various generative architectures and models, but does not clearly specify which real-image datasets were used in the spectral analysis. Could the authors clarify the datasets used and provide empirical evidence or a systematic explanation of whether Spectral Tail Uplift is consistently absent across diverse real image datasets?
  2. Could the authors explain why the proposed representation alignment is more effective than simpler feature integration approaches, such as weighted fusion of spatial and frequency features? Additionally, how are spatial and frequency features combined in the Direct Spatial-Frequency Inference baseline in Table 2? An ablation isolating the contribution of the alignment loss would also be helpful.
  3. What motivates the selection of (ρ∈[0.7,1.0]\rho\in[0.7,1.0]) as the extreme high-frequency range? Could the authors provide a clearer justification or a brief experimental comparison using alternative frequency ranges?
  4. The submission includes neither implementation code nor a Reproducibility Statement. Although both are optional, they are strongly recommended, as reproducibility is crucial for research reliability and future progress. Do the authors plan to release the code and training configurations necessary to reproduce the reported results?

I appreciate the authors' efforts to advance research in AI-generated image detection. I would be willing to reconsider and potentially increase my score if the questions and weaknesses mentioned above are adequately addressed, particularly the generalizability of the absence of Spectral Tail Uplift across diverse real-image datasets and the concerns regarding reproducibility.

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

Related papers

Discussion (0)

Sign in to comment.