acceptodds
Withdrawn from ICLR 2027

The Tell-Tale Denoiser: Diffusion Model Attribution via Spectral Geometry Signatures

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Attributing a generated image to its source diffusion model is a fundamental challenge in provenance verification and intellectual property protection. This problem is particularly difficult because diffusion models trained on different datasets can converge to similar score functions and thus similar output distributions, making the generated images themselves unreliable as attribution evidence. Existing non-invasive methods either fail on architecturally similar variants or rely on signals that vanish when models share the same autoencoder. We propose Spectral Geometry Signatures (SGS), a non-invasive attribution method that fingerprints each candidate model's denoising behavior rather than its outputs. Our key insight is that a model's score function exhibits a distinctive spectral geometry (how it redistributes energy across spatial frequency bands during denoising) that is intrinsic to the denoiser and invariant to the choice of autoencoder. SGS extracts this signature via frequency-controlled perturbations and standard forward passes, with no inversion, optimization, or generation-time enrollment. Across closely related diffusion checkpoints and models differing in architecture and training procedure, SGS provides highly discriminative attribution signals, remains effective under cross-domain prompt shift, and is stable across the autoencoder choices we evaluate. Our code is available at https://anonymous.4open.science/r/SGS-1B08.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Fingerprinting, Attribution

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_o2cfOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Spectral Geometry Signatures (SGS) for attributing generated images to their source diffusion models. SGS probes candidate denoisers with frequency-controlled perturbations and measures their cross-frequency responses, without model inversion or modification. Experiments cover diffusion and flow-matching models, achieving 99.99% and 99.98% accuracy in 15-model and 19-model closed-set settings, respectively. The authors also evaluate robustness to prompt shifts, image distortions, and different VAEs.

Major Comment

This paper presents an interesting approach and reports impressive results across several settings. However, I have two main concerns: whether the near-perfect accuracy reflects genuine image-source attribution rather than candidate-model identification, and whether the limited baseline comparisons sufficiently establish the method's advantage over recent work.

These issues are important for assessing the paper's central claims. I lean toward rejection, but would reconsider if the authors provide convincing cross-model attribution analysis and stronger comparisons.

Critical strengths

1. Interesting approach to model attribution. Instead of relying on image artifacts or reconstruction errors, SGS uses the spectral responses of denoising networks. This offers a different perspective on model attribution, particularly for closely related models such as SD v1.4 and SD v1.5.

2. Strong experimental results. SGS achieves near-perfect closed-set attribution accuracy across different model architectures and fine-tuned variants. The experiments also cover flow-matching models, prompt distribution shifts, image distortions, and VAE variations.

Critical weaknesses

1. Model fingerprints are not clearly distinguished from image source attribution.

SGS extracts features by directly querying each candidate denoiser, which may naturally reveal the model's identity. However, distinguishing denoising networks does not necessarily establish which network generated a given image.

This concern is amplified by the large gap between the training-free Argmin method (78.33% on 8 models and 65.56% on 15 models) and LinearSVC (approximately 99.98% and 99.99%, respectively). The paper does not sufficiently explain what the supervised classifier learns or how much of its performance comes from image-source matching rather than candidate-model identity.

A cross-model analysis, where images from different sources are evaluated using the same candidate denoisers, would help clarify this issue.

2. Insufficient comparison with recent attribution methods.

The main comparison includes only RONAN and LatentTracer, both based on inversion or reconstruction-related signals. This is insufficient to establish the advantage of SGS over other attribution approaches.

Relevant studies include Detecting Origin Attribution for Text-to-Image Diffusion Models, Who Made This? (FRIDA), and Tracing the Roots. Although these methods have different assumptions, comparisons under matched settings, where applicable, would better establish the contribution of SGS.

Questions for authors

Q1. Can the authors provide a cross-model attribution score matrix, showing how each candidate denoiser responds to images generated by different source models? Including unrelated or real images as controls would help verify that SGS captures image-source information rather than only model identity.

Q2. Why does LinearSVC outperform Argmin by such a large margin? Can the authors clarify what information the classifier learns and whether its predictions depend on the true image source?

Q3. The full SGS configuration requires 1,200 denoiser forward passes per candidate model. Can the authors report inference time and computational cost against existing methods? Also, since a single timestep already achieves 99.97% accuracy on 15 models, what practical benefit justifies the full five-timestep configuration?

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

This paper studies source attribution for diffusion-generated images, i.e., identifying a given image is produced by which model in a closed set of candidates. The authors argue that model identity mainly lies in the behavior of the denoiser and propose SGS, which injects controlled noise into different frequency bands of the latent and measures how the denoiser's prediction error distributes its energy across them. This distribution serves as the signature of each model, and attribution is performed with prototype matching or a linear classifier. The authors evaluate SGS on up to 19 models, covering same-family fine-tunes, different architectures, distilled models, and flow-matching models, and report near-perfect attribution accuracy (close to 100%) in many settings.

Critical strengths
  • The problem is practical and relevant. Same-family fine-tuned models are abundant in the open-source ecosystem, and attribution under a shared VAE is indeed challenging. The authors show that decoder-inversion methods fail structurally when candidates share a VAE, a valuable observation which provides a reasonable motivation for the work.
  • The method requires only forward passes and leaves the models unchanged, so it is simple to implement. The additional experiments cover a broad range of settings, and it is clear that the authors put considerable effort into validation.
Critical weaknesses
  • Important related work is missing, and the comparison with Wang et al. [4, 5] might be unfair.
    • Related work:
      • FLIPAD [1] has already studied the same problem on SD versions that share the same autoencoder, such as SD v1.1 and v1.4, but it is not cited or compared
      • Floros et al. [2], the closest work in terms of mechanism, also performs white-box model attribution using denoiser responses across multiple timesteps. The paper did not compare with [2].
      • AEDR [3], a more recent training-free method from the same family as the baselines, is also not discussed.
    • Comparison: the comparison with RONAN [4] and LatentTracer [5] in Table 4 has three issues.
      • Both methods were originally designed for single-model verification, i.e., deciding whether an image was generated by a given model. The paper does not explain how they were adapted to closed-set n-way attribution setting.
      • The authors of LatentTracer [5] have pointed out that its signal comes from the decoder. Consequently, its failure on the shared-VAE subset is natural, and the gap on this subset provides limited evidence for the advantage of SGS.
  • I suspect that the evaluation may be affected by data leakage, for the reasons detailed below.
    • First, several details in the paper are inconsistent with the identification protocol described in Sec. 4.2.
      • As stated in Sec. 4.2 and Eq. 9, the source of a query image is unknown. Therefore, its signature Φ(Mi,x)\Phi(M_i, x) should be computed with every candidate MiM_i. In Fig. 1, however, the query signature is written as ϕ(M∗,z0)\phi(M^{*}, z_0), i.e., it is computed with the true source model M∗M^*.
      • In Table 3, Gaussian blur with σ=8 largely removes the image content, yet the accuracy of LinearSVC on MS-COCO rises from 96.20% to 99.9%.
      • In Table 11, argmin achieves only 47.2% and 43.6% on SD v1.4 vs. v1.5, which is below random chance, while LinearSVC exceeds 98% under the same setting.
    • The provided code (eval_all8_1t_5t.py) supports this suspicion.
      • In run_plan (line 719), the signature of each test image is computed with wrappers[mkey], where mkey is the true source of that image. In other words, each test image is probed with the denoiser of the model that generated it, which is impossible at deployment time because the source is unknown.
      • Both argmin and all classifiers operate on these signatures, and the per-candidate probing protocol of Eq. 9 is not implemented in the code.
      • The signature cache path depends only on the model name, the prompt index, and the split. Consequently, a cross-probed signature Φ(Mi,xMj)\Phi(M_i, x_{M_j}) would be written to the same file as Φ(Mi,xMi)\Phi(M_i, x_{M_i}), which indicates that the code does not support cross-probing.
    • Therefore, I strongly suspect that the accuracy of about 99.9% in Table 2 results from label leakage.
      • The paper argues in Sec. 3 and Appendix C that the signature mainly reflects intrinsic properties of the denoiser and depends only weakly on image content. Since the test signatures are computed with the true source model, the denoiser-induced differences in the signatures directly encode the label. A classifier can answer correctly by recognizing which denoiser produced the signature, without using any source information contained in the image.
      • This explanation is consistent with the accuracy increase after blurring, as well as with the large gap between argmin and the linear classifiers, and this is the main reason for my soundness score of 1.

I believe a rigorous check and clarification is needed for the paper to be accepted, and I'm happy to be corrected if I misunderstood anything.

References:

[1]: Laszkiewicz et al., Single-Model Attribution of Generative Models Through Final-Layer Inversion. ICML 2024

[2]: Floros et al., Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution. NeurIPS 2025

[3]: Wang et al., AEDR: Training-Free AI-Generated Image Attribution via Autoencoder Double-Reconstruction. AAAI 2026

[4]: Wang et al., Where Did I Come From? Origin Attribution of AI-Generated Images. NeurIPS 2023

[5]: Wang et al., How to Trace Latent Generative Model Generated Images without Artificial Watermark? ICML 2024

Questions for authors
  1. Please explain the following results:
  • Argmin performs poorly on the key shared-VAE models. In the 15-model registry of Table 8, SD v1.5 reaches only 39.6% and Dreamshaper-8 only 30.8%. In the SD v1.4 vs. v1.5 binary task of Table 11, argmin achieves only 47.2% and 43.6% overall, which is below random guessing, while LinearSVC exceeds 98% under the same setting. What accounts for such a large gap?
  • In Table 3, Gaussian blur with σ=8 removes almost all content of a 512-resolution image, yet the accuracy of LinearSVC on MS-COCO increases from 96.20% (Table 2) to 99.9%. The authors attribute this to the claim that "isotropic noise that blurs individual-sample distances but leaves class centroids intact", but blurring is not isotropic noise. Does the accuracy increase after the content is destroyed indicate that the classifier does not rely on image content?
  • Rotation by 75 degrees, 75% cropping, and JPEG compression with Q=25 also have almost no effect on the classifier. Why is that?
  • Please provide code implementing the protocol of Eq. 9, i.e., computing the signatures of each query image under all candidates and including mismatched samples such as Φ(Mi,xMj)\Phi(M_i, x_{M_j}) when training the classifier. Please also report Table 2, Table 4, and Fig. 5 under this protocol. In addition, if the same image is probed by every candidate model and the resulting signatures are fed to the trained classifier, does the prediction change with the probing model?
  1. Under their original protocols, RONAN [4] and LatentTracer [5] are single-model verification methods, i.e., they decide whether an image was generated by a given model. How were they reproduced for closed-set n-way attribution, and how were the thresholds and scores calibrated? Does the SGS column in Table 4 use argmin or LinearSVC?
  2. In an open-source ecosystem where LoRA and model merging are widespread, how realistic is the closed-set assumption? Table 12 shows that open-set rejection remains weak for same-family fine-tunes. How do the authors view the impact of this limitation on practical deployment?
Reviewer_N8azOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper asks whether we can identify which diffusion model generated an image, given access to a collection of possible source models. Output-based classifiers struggle with closely related model variants, and inversion-based methods are slow and lose signal when models share an autoencoder. To address this, the authors propose Spectral Geometry Signatures (SGS). SGS encodes an image to a latent with a fixed VAE, adds noise restricted to one of K radial frequency bands, runs one denoiser forward pass, and measures how the residual energy (predicted minus injected noise) is distributed across the K bands. Repeating this over the different bands, amplitudes, and timesteps gives a 960-dimensional feature vector which the authors call a "signature." At query time, the image is passed through every candidate model's denoiser. The method uses either average reference vectors or a trained classifier to predict the image’s source. The paper reports near-perfect LinearSVC accuracy on 8 and 15 model comparisons, robustness to prompt shift and image distortions, large gains over the inversion based methods like RONAN and LatentTracer, and extensions to flow-matching models.

Critical strengths
  1. Probing the denoiser is a reasonable direction. Sec. 3 shows that output-space features cannot separate SD v1.4 from v1.5. The paper also correctly points out that inversion baselines lose their signal when models share a VAE. Looking at the denoising behaviour rather than outputs is well motivated, and a forward-pass-only method would be useful in practice.
  2. Broad model testing sweep. The tested models covers same-architecture fine-tunes, different architectures, a distilled model, and flow-matching models. This would be a strong stress test if the evaluation protocol is sound.
Critical weaknesses
  1. The paper’s explanation of the method and its reported results are inconsistent, and the main numbers may not measure image attribution. The paper repeatedly describes the spectral feature vector as intrinsic to the model and largely independent of image content. It also states that at query time every candidate model's denoiser is probed and “the source model is unknown at query time and is never assumed” (Sec. 4.2, line 323). These two claims together imply that every candidate denoiser's spectral feature vector on any query should resemble that candidate’s own prototypes. The selection step would then carry little information about the image. If image attribution works, the signature must depend on whether the image came from the probing model, but the paper is not clear that it states or tests this mechanism:
  • The training-free rule that follows Eq. 9 reaches 78.33% on 8 models (Table 2, SD-Prompts), and 53–56% in-domain on SD v1.4 vs v1.5 (Sec. 6.2 line 426), which is close to chance on the SD v1.4/v1.5 pair the paper uses to motivate the method.
  • Under a distortion that removes most image content (σ=8\sigma=8 Gaussian blur on MS-COCO (Table 3)), the training-free rule (Eq. 9) drops from 78.4% to 21.3%, while LinearSVC rises from 96.2% to 99.9%. The rule that actually compares each candidate’s response to its prototype fails once the image is degraded, but the classifier improves. The paper instead attributes this to the fingerprint being something pixel-space distortions “cannot erase regardless of strength” (lines 446–447). For an image attribution method, accuracy that is insensitive to the image may be a concern.
  • Sec. 5 describes LinearSVC as trained with 5-fold cross-validation on the denoiser spectral feature vector. The paper does not say which denoiser computed the training and test spectral feature vectors.

These results are what one would expect if each test spectral feature vector were computed with the image’s true source denoiser. In that case, the classifier would only need to recognize which denoiser produced the signature. Under the white-box access SGS assumes, that task may be trivial, since one could compare weights directly. Until the protocol is clarified, I cannot tell whether the headline results measure image attribution or identification of the probing denoiser.

  1. Although SGS avoids iterative optimization, it makes many denoiser queries. The stated settings require 1,200 evaluations to obtain one image’s feature vector using one model. Testing 8 possible source models requires 9,600 evaluations per image. The paper should report the total time and memory required to attribute an image, alongside comparable measurements for the baselines.
Questions for authors
  1. For each reported result, which denoiser computes the signatures the classifier is trained and tested on?
  2. If images generated by model A are probed through model B’s denoiser, what does the classifier predict? If it predicts B, the spectral feature vector encodes the probing model rather than the image’s source.
  3. On real images, lets say Imagenet, generated by no candidate, are probed through each candidate’s denoiser, does SGS confidently attribute them to some model?
  4. What mechanism makes a model’s signature on its own images closer to its prototype than its signature on other models’ images? Given that the signature is described as content-independent, where does the image-specific signal come from?
  5. Table 3 reports 99.9% accuracy on 160 test images per distortion. With 160 images, the achievable values near 100% are 99.4% and 100%. How was this number computed?
  6. How is SGS applied to the flow-matching models? The paper states that all images are encoded using the SD v1.5 VAE, but I am not aware that the evaluated flow-matching models share a compatible latent space. Were their own encoders used instead, and how were the resulting spectral features combined in the 19-model evaluation? Also, how is the residual defined for models that predict velocity rather than noise?

Other notes:

  • The paper would be much easier to follow if Sec. 3 and Sec. 5 stated explicitly which denoiser computes each signature in every experiment.
  • Typo: double period at line 239.

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

Related papers

Discussion (0)

Sign in to comment.