acceptodds
Under review as a conference paper at ICLR 2027

PIXEL ECHO: MULTISCALE PHASE DECOMPOSITION OF GENERATOR FINGERPRINTS

Abstract

Generated images carry recurring traces locked to the pixel grid. Where these traces originate decides what an image can be attributed to and which operations erase them. We introduce Pixel Echo, a diagnostic measurement that resolves these traces by spatial phase and repetition period. Pixel Echo pools local pixel responses at matching grid positions and exactly decomposes the resulting periodic structure into count-weighted orthogonal increments at periods of 2, 4, 8, and16 pixels. Their strengths define a scale profile for examining how fingerprints vary across decoders and input routes. Controlled comparisons across two decoder families show that reconstructions and generations sharing a decoder exhibit closely aligned patterns, while pixel-head replay links one decoder’s pattern to phase-specific projection weights. Fixed-latent decoder swaps reveal a perioddependent distinction: decoder-associated patterns concentrate at periods up to the upsampling factor, whereas longer-period structure can pass through from the input. Specifically, six ×8 decoders differ at periods 2–8 but retain a 16-pixel component shared with each other and with real photographs, whereas ×16 decoders exhibit decoder-specific patterns at this period. On held-out splits of our class-to-image and text-to-image benchmarks, linear classifiers frozen beforehand achieve 95.1% four-way and 90.9% seven-way source attribution accuracy, increasing to 98.6% and 94.9% when combined with frequency features, at 0–1.4% per-source false-positive rates on real images. These results support closed-set attribution of decoder and export pipelines rather than unique identification of upstream generators, with attribution requiring preservation of the pixel grid. Code, benchmark manifests, frozen experiment definitions and an interactive demo are available at https://pixel-echo.github.io/.

Primary area:
generative models
Keywords:
Generative Model, Fingerprints, GenAI 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_Yok7OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Pixel Echo, a fixed statistic that pools the luminance Laplacian response into a 16x16 phase grid and exactly splits it into four period increments (P2/P4/P8/P16). They show generated images and the same decoder's reconstructions share a strongly aligned fingerprint pattern (r ~ 0.996) but differ in amplitude, and the decoder's own structure only appears at periods up to its upsampling factor. On a self-built closed-set benchmark, a 340-d feature with frozen linear heads reaches 95.1%/90.9% source attribution for C2I/T2I, and 98.6%/94.9% when frequency features are added.

Critical strengths

1、The paper cleanly shows generated images carry the same fingerprint as their decoder's own reconstructions, and that this fingerprint only exists at periods up to the upsampling factor — which also explains why some sources are much harder to detect than others. 2、 The experiments are unusually careful and honest: the classifiers are frozen before testing, the statistics are properly corrected, and the authors also report results that are bad for them, like their method failing after one resize.

Critical weaknesses

1、The method is a hand-crafted statistic with no learned part, and it is close to earlier work on up-sampling artifacts. The key finding is only checked on about 8 decoders at 256x256 ImageNet, so we do not know if the period rule holds for other resolutions or longer downsampling chains. 2、The method breaks easily in real use. One ordinary resize drops attribution to chance, JPEG compression halves it, and almost all shared images online have passed such steps — so it mostly works only on original untouched exports. 3、Detection is weak on the hardest and most common sources: DiT is caught only 1.6% of the time, and same-decoder sources are barely above guessing. The best numbers actually come from adding the older frequency features, not from Pixel Echo itself.

Questions for authors

1、Beyond the 256x256 ImageNet, how much of the periodic rule is retained - do you think that for decoders running with other output sizes or multi-stage upsampling chains, the writing methods of P16 and P8 will change? 2、Does your real dataset guarantee that it has not undergone upsampling or re-encoding? A "real" dataset contaminated by the decoder would expand the shared P16 parameters in Section 4.3. 3、What is the joint full-head false positive rate in the real situation when all sources' heads are triggered simultaneously? Does this exceed the 0.8% to 1.4% per head that you reported? 4、Can you quantify that 98.6% of Max's attribution comes from only the additional frequency features, such as using only the frequency heads in the same test? 5、After bilinear resampling, is there a robustness (such as phase grid estimation) that can restore a higher probability attribution? Or is the grid requirement fundamental? 6、ORT is cited only as a 2026 GitHub repository. Can you confirm its checkpoint, prompts and export settings are frozen and verifiable, given the Test is a single pass?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Yok7 (1/2): W1, W2, W3

We thank the reviewer for the careful reading and for recognising the frozen-before-test protocol and the reporting of unfavourable results. Two statements in the review are factually incorrect (W1 on scale, W3 on the frequency features); we correct them first, with the evidence.

W1 (hand-crafted; ~8 decoders at 256²). Correction: the period rule is tested on twelve decoders from 256² to 1024², not on ~8 decoders at 256² only. The operator audit (Table 20) covers seven ×8 decoders (four kl-f8 weight sets, the SDXL VAE, TAESD, the consistency decoder), three ×16 (AliTok, RAE, VAR) and two pixel-space heads (JiT, PixelDiT). Beyond 256²: six ×8 decoders at 512² (Table 17), seven disjoint 7,000-image cohorts at 512² (Table 21), a 384² latent crop (P16 r = 0.93 to its 512² decode) and SDXL at 1024² (Table 18). The T2I benchmark adds the strongest commercial services available before submission (GPT Image 1/2, Imagen 3, Nano Banana, Seedream 3) at their native resolution. Only the ×16 evidence is at 256². The statistic is hand-crafted by design, so that each period can be read separately; §3.2 now states its DFT equivalence and §2 adds AEROBLADE and Sundara Rajan et al. (2025).

W2 (breaks under ordinary processing). The paper states this scope up front — abstract, §1, §5.3 and the first item of Appendix I with Table 37 — and the revision now removes its translation part. The folded-map DFT magnitude keeps 95.0 % C2I attribution under a random crop-shift, where the published feature gives 49.6 % (Table 38); resampling destroys the structure itself (Q5). One correction: JPEG q90 does not halve C2I attribution (81.5 %); it does on T2I (40.4 %), and at q75 C2I keeps 57 %. The method targets raster-preserving media such as native service exports, where held-out T2I attribution is 90.9 % across seven services.

W3 (weak detection; best numbers from frequency features). Correction: on T2I the best numbers come from Pixel Echo, not from the frequency features. On the held-out T2I test Pixel Echo alone reaches 90.9 % against 83.1 % for Frequency-64D, and inside Max the Pixel Echo block carries 9.5 pp against 3.8 for the frequency block (Table 31); on C2I the frequency features are indeed strong (95.3 vs 95.1). Low DiT detection at 1 % FPR is what the period rule predicts for a ×8 decoder, yet attribution recognises DiT at ≥ 91.2 % recall (Table 4). That sources sharing one decoder are hard to separate is the finding that the signal identifies the decoder (Table 32), not a failure of attribution.

AuthorsOpenReview ↗

Response to Reviewer Yok7 (2/2): Q1–Q6

Q1 (other resolutions, longer chains). For ×8 decoders the rule holds from 256² to 1024², and the commercial services match its prediction. A ×8 system should carry no decoder-specific P16: Imagen 3, Nano Banana and Seedream 3 put ≤ 1 % of their energy there (Table 18). The ×16 decoders are ImageNet-256 models, so ×16 is measured at 256².

Q2 (is the Real data free of resampling?). Yes for the cohorts behind our claims — and we found and report exactly this contamination elsewhere. Benchmark ImageNet Real is cropped without resizing; the matched cohort's per-image bicubic crop shows no phase-locked P16 (split-half −0.13). COCO Real, resampled by 16/15, does (1.00) and is analysed in Appendix E.5.

Q3 (joint false-positive rate). 2.6 % for Pixel Echo, 3.6 % for Max with all four heads together (Table 35) — above the per-head rates, as a union must be.

Q4 (Max without Pixel Echo). On the same held-out test, Frequency-64D alone gives 95.3 % (C2I) and 83.1 % (T2I) against 98.6 % and 94.9 % for Max: Pixel Echo adds 3.3 and 11.8 pp (Tables 4–5).

Q5 (phase-grid estimation after resampling). It restores translation, not resampling. A 256-offset search recovers a crop-shift (90.6 %), but after bicubic rescaling every variant is at chance (Table 38): the period-16 lattice no longer exists.

Q6 (is ORT frozen and verifiable?). Yes. Public weights (https://huggingface.co/donglixu/ORT) and paper (https://openreview.net/forum?id=s8kdXG7Zu4); chosen as the strongest published latent autoregressive generator on a conventional tokenizer we could find. Checkpoint, conditions, seeds and export settings were fixed before the test (Appendix H.1).

Thank the authors for the rebuttal and accept the corrections on decoder scale, JPEG and the T2I ablation. However, my concerns about practical scope, generality and novelty remain, so I keep my rating.

AuthorsOpenReview ↗

The accepted corrections were the basis of the remaining concerns

Thank you for the reply. We note that the three corrections you accept are the evidence the review gave for these concerns:

  • Generality rested on "about 8 decoders at 256x256". It is twelve decoders, 256² to 1024².
  • Practical scope rested on JPEG halving C2I attribution. It does not: 81.5 % at q90.
  • Detection rested on frequency features carrying the best numbers. Combining the two adds 3.3 points on C2I and 11.8 on T2I over frequency features alone (98.6 and 94.9), so they carry different information. K5hR (S3): "phase-aligned structure carries information that spectral magnitude features alone do not"; vb5L: the improvement from combining them "demonstrates the value of the proposed representation". On T2I the best numbers come from Pixel Echo (90.9 vs 83.1).

Novelty. Our finding is new: a decoder leaves its own pattern only at periods up to its upsampling factor. Our analysis is new: we tested this with hypotheses fixed before seeing the data. Our application is new: using this pattern, Pixel Echo gives the best T2I attribution of all methods we compare (90.9 %), on the newest commercial services at their native high resolution. Our benchmark is new: class-to-image and text-to-image sets built from the latest open models and commercial services, with manifests released with the code. If earlier work already shows any of this, please name it and we will compare.

If a further reason stands behind any of the three concerns, please let us know which one. The same holds for the other reviews: if you see a point our replies there leave open, tell us. We stand by those answers, and we can run any analysis today and post it before the discussion closes.

Reviewer_vb5LOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method to analyze generator fingerprints using spatial phase patterns and an exact multiscale decomposition at periods of 2, 4, 8, and 16 pixels. The authors investigate how these patterns relate to decoder architectures and input latents using reconstruction comparisons, fixed-latent decoder swaps, and pixel-head experiments. They find that images sharing a decoder exhibit similar fingerprints, while the dominant fingerprint scales depend on the decoder's upsampling structure. On held-out benchmarks, Pixel Echo achieves 95.1% four-way and 90.9% seven-way source attribution accuracy, improving to 98.6% and 94.9% when combined with frequency features.

Critical strengths
  1. The controlled decoder swaps and reconstruction experiments provide convincing evidence that fingerprint patterns are strongly associated with decoder structure. The multiscale analysis offers useful insights beyond conventional fingerprint-based classification.
  2. The paper evaluates multiple decoder families and generation pipelines, includes several competitive baselines, and reports held-out results. The improvement from combining Pixel Echo with frequency features also demonstrates the value of the proposed representation.
Critical weaknesses
  1. The core method combines a fixed Laplacian filter, phase-conditioned averaging, and orthogonal projections onto nested periodic subspaces. These are established signal-processing operations, and the decomposition is an invertible reparameterization of the original phase map. The paper does not sufficiently demonstrate why this particular decomposition provides a substantial methodological advance beyond existing periodic or spectral fingerprint analysis.
  2. The method relies heavily on preserving the original pixel grid. As shown in Table 34, a single bicubic resampling reduces C2I attribution accuracy from 95.7% to approximately 23%, near chance. This significantly limits its usefulness for real-world image forensics, where resizing and recompression are common
Questions for authors
  1. Since the full multiscale representation contains the same information as the centered phase map, what is the main advantage of the decomposition beyond interpretability?
  2. Is the severe degradation under resizing primarily caused by phase misalignment or by destruction of the underlying periodic signal?
  3. How well would Pixel Echo distinguish unseen decoder checkpoints or generators sharing similar decoder architectures? This seems important for establishing whether the fingerprints capture generalizable decoder properties rather than specific benchmark pipelines
1 comment
AuthorsOpenReview ↗

Response to Reviewer vb5L

We thank the reviewer for a clear reading of the paper and for questions that directly improved the revision — Q2 in particular became a new experiment (Table 38).

W1 (established operations). The operations are indeed standard, and we now say so explicitly; the contribution is what they make visible. §3.2 states that the increments are a dyadic grouping of the folded map's 2-D DFT coefficients (Proposition 1, Table 30), and §2 adds AEROBLADE and Sundara Rajan et al. (2025). Read period by period, the map shows that a decoder writes its own pattern only up to its upsampling factor, while longer periods carry input structure shared across decoders: six ×8 decoders on disjoint latent batches still agree at P16 (0.69–0.83, Table 21). A whole-field or spectral summary would hide this (0.12 overall vs 0.82 at P16).

W2 (needs the pixel grid). A fair concern, and one the paper states up front — in the abstract, §1, §5.3 and as the first item of Appendix I, where Table 37 quantifies seven transforms. The revision now separates its two parts: translation is handled, resampling is not. The folded-map DFT magnitude keeps 95.0 % C2I attribution under a random crop-shift, where the published feature gives 49.6 % (Table 38). Resampling sensitivity is the other side of what makes the signal specific: it is locked to the decoder's output grid. The method therefore targets raster-preserving settings such as native service exports and provenance checks at the point of generation, where it reaches 90.9 % held-out on seven T2I services; a resampling-aware variant is a natural next step.

Q1 (advantage beyond interpretability). Its value is localisation: it tells which period carries a difference, and that is what makes the decoder/input split above observable. For classification it adds no information, as Proposition 1 and Table 30 already stated. Decoder swaps change P2–P8 and leave P16; without the boundary the whole-field correlation with Real is 0.06 while the P16 increment alone is 0.48–0.64.

Q2 (misalignment or destruction?). Destruction — and thank you for the question, which we turned into an experiment. Phase alignment recovers a crop-shift (DFT magnitude 95.0 %, offset search 90.6 %), but after bicubic ×0.9/×1.1 every variant is at chance (≤ 24.7 % vs 25 %): a non-integer rescale moves the period-16 lattice to non-integer periods (Table 38).

Q3 (unseen checkpoints, shared decoders). The signal identifies the decoder checkpoint rather than the generator, which we see as the natural granularity of what the measurement captures. Two decoders of one architecture are separated at 86.6 % (SD1.5 original kl-f8 vs ft-mse), generators sharing one decoder only 9–11 pp above chance, and family labels reach 77.4 % against 54.8 % for generator labels (Table 32). Transfer to checkpoints unseen in training is listed as not evaluated in Appendix I ("What is attributed") and is an important next step.

Reviewer_K5hROpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Pixel Echo, a multiscale phase-based representation for analyzing periodic fingerprints in AI-generated images. The authors investigate the relationship between these fingerprints and decoder architectures through reconstruction-generation comparisons, decoder swapping, and pixel-head analysis. The method achieves 95.1% and 90.9% accuracy in closed-set C2I and T2I source attribution, respectively, with further improvements when combined with frequency features. However, its performance is sensitive to common image transformations.

Critical strengths

S1. A clean shared-decoder comparison. Matching real-image reconstructions against generations through the same AliTok decoder isolates the decoder as the common factor. The resulting alignment (r = 0.996, against a median of 0.06 across different decoders) is clear, and a second family based on sd-vae-ft-mse is included, albeit with weaker alignment (0.846).

S2. A rigorous protocol and candid reporting. Heads and thresholds are frozen before a single-pass held-out test, and predictions for the fixed-latent series were committed in advance. Failed predictions are reported rather than dropped. The metadata-only shortcut audit, boundary controls, convergence sensitivity analysis, and robustness table make the scope of each claim unusually easy to assess.

S3. Complementarity with frequency features. Appending Pixel Echo to frequency features improves attribution on both benchmarks and under all three optimizers. This suggests that phase-aligned structure carries information that spectral magnitude features alone do not.

Critical weaknesses

W1. Limited novelty of the measurement and missing related work. For images whose sides are multiples of 16, the count weighting is uniform, and the nested increments become exactly a dyadic grouping of the 2-D DFT coefficients of the folded map. Folding the residual modulo 16 is equivalent to sampling its image-averaged complex spectrum at multiples of 1/16 cycles per pixel, so the scale profile is the spectral energy of the averaged residual within these groups. The measurement is therefore closely related to averaged residual fingerprints and their spectra (Marra et al., 2019; Corvi et al., 2023), and Proposition 1 and Table 28 confirm that the decomposition adds no information for classification. The finding that reconstructions and generations share latent-decoder traces is also established in work the paper does not cite, notably AEROBLADE (Ricker et al., 2024) and training on aligned autoencoder reconstructions (Rajan et al., 2025).

W2. The P16 findings stated in the abstract are confounded. In the fixed-latent series, all six decoders decode the same 1,000 latents, so their P16 increments share the finite-sample content component of the same images. Since the interior P16 split-half reliability of sd-vae-ft-mse is only 0.38 (Table 22), cross-decoder agreement in this paired design cannot be attributed to input structure without a disjoint-batch control. The latent-roll experiment does not resolve this, because content-driven variation also moves with the content. The claim that this component is shared with real photographs rests on full-raster correlations. Yet the replicate-padding ring falls on phases 0 and 15 for every image whose side is a multiple of 16 and projects mainly onto P16, and the interior-only analysis lowers the Real ImageNet correlation of the sd-vae family from 0.46 to 0.06 (Fig. 13). The corresponding interior P16-increment value is not reported. The bound itself also follows from shift equivariance (Appendix E.6), and the fixed-latent predictions failed in three of four rounds, yet the abstract and introduction present the period rule as an established finding.

W3. The mechanism evidence is largely predetermined by the architecture. In AliTok’s pixel head, each token’s hidden state is shared by all 256 pixels of its block, and the only position-dependent parameters are the per-position weights and biases. Any within-block phase pattern must therefore come from them. Replacing the weights with their phase average makes each block constant, which explains the border-dominated map in Fig. 5. The replay confirms the architecture rather than localizing a mechanism, and the single output-matched intervention showed no scale selectivity. The scale profile is also weakly informative. AliTok’s shares match the degrees-of-freedom split of a structureless template, as the authors note, and Real ImageNet’s profile is similar, yet no noise-floor correction is applied to cohorts whose template amplitude is close to that of real images.

W4. The application results are weaker than presented. On the held-out C2I test, Pixel Echo alone matches Frequency-64D in four-way attribution (95.1% vs. 95.3%) and in average detection recall (55.5% vs. 54.3%). At an L-BFGS stationary point, the NPR probe surpasses it (97.5% vs. 95.6% Validation macro-F1). The held-out tables use heads that mostly did not converge, only 18 of 67, which disproportionately disadvantages high-dimensional probes. The detection columns also compare per-source heads trained on each target generator against zero-shot official detectors, and joint detection over unknown sources is not evaluated. Because the SD-family templates have amplitudes comparable to or below that of real images, detection is effectively limited to the ×16 linear-head decoders, with 1.6% recall on DiT. Together with the collapse under the transformations listed above, this leaves the practical scope narrow.

W5. Presentation. The main text relies on internal registry labels such as E29 and C14, as well as file paths and a truncated claim registry, and it contains revision traces such as “Reviewers asked whether…” (H.4). Key evidence is spread across 39 pages, which makes it hard to verify the main claims.

Questions for authors

Q1. Could the authors repeat the fixed-latent comparison using disjoint latent batches for each decoder, so that cross-decoder P16 agreement is measured without shared image content? If the agreement then falls toward the split-half floor, the P16 sharing would reflect content sampling rather than structure that passes through the decoder. I would also ask the authors to report the interior-only P16-increment correlation with Real ImageNet, since the full-raster value underlies the claim in the abstract. This is the question most likely to change my assessment of the mechanism claims.

Q2. How do the scale profiles compare with a noise-floor baseline estimated at the same sample size, for instance from split-half cross-energy? Several profiles in Fig. 6 are close to the degrees-of-freedom split, so without such a correction it is unclear which differences reflect decoder structure.

Q3. Since the held-out test relies mostly on non-converged heads, could the authors report the main attribution results with all heads optimized to stationarity? It would also help to compare Pixel Echo Max against a concatenation of baseline features, such as Frequency-64D combined with the NPR probe. That comparison would show whether the fusion gain is specific to phase-aligned features or simply reflects combining two complementary representations.

Q4. In the T2I benchmark, Seedream’s lower-period energy is roughly two orders of magnitude above that of GPT Image 1 (Table 7), which suggests a post-processing or upscaling stage. Some closed services also embed invisible pixel-level watermarks. How much of the T2I attribution reflects such pipeline stages or watermarks rather than passive decoder fingerprints? For open-weight models, comparing locally decoded outputs with those from a hosted service would help separate the decoder from the service pipeline.

Q5. Could the authors evaluate a shift-invariant variant, such as the DFT magnitudes of the folded map or an exhaustive search over the 256 phase offsets? An unknown crop offset currently reduces C2I attribution to about 50%. Such a variant would show how much of the method’s advantage depends on retaining the phase origin and how much remains when the feature reduces to a spectral statistic.

2 comments
AuthorsOpenReview ↗

Response to Reviewer K5hR (1/2): W1, W2, W3

W1 (novelty, related work). Changes: §3.2 now states that for rasters whose sides are multiples of 16 the increments are a dyadic grouping of the folded map's 2-D DFT coefficients and add no information to a classifier (Proposition 1, Table 30); §2 adds AEROBLADE (Ricker et al., 2024) and Sundara Rajan et al. (2025). The novelty is not the transform but what reading it period by period reveals: a decoder's own pattern lives at periods up to its upsampling factor, while longer periods carry input structure shared across decoders (W2). A single field-level or spectral number hides this: two decoders of the same latents agree at 0.12 over the whole field but 0.82 at P16, and without the boundary the whole-field correlation with Real falls to 0.06 while the P16 increment stays at 0.48–0.64. The same reading gives same-decoder alignment across input routes (0.996 AliTok, 0.846 sd-vae-ft-mse; 0.06 across decoders).

W2 (the P16 statement in the abstract). Each point has a specific cause; the new control resolves it.

(i) Reliability 0.38, shared latents. The ×8 P16 signal is weak (template amplitude ~20× below AliTok, Table 20), so at n = 1,000 the template is mostly sampling noise, which decoders fed the same latents share. With disjoint 7,000-prompt batches per decoder (Table 21, pre-registered), the six decoders still agree at 0.69–0.83 (median 0.78), above two disjoint batches of the same decoder (0.67).

(ii) Latent roll. Rolling the latent also moves the content, so it cannot separate the two explanations; (i) does.

(iii) Real correlation 0.46 → 0.06 (Fig. 13). The padded ring is shared by all images, and that number is for the whole field, whose ×8 energy lies mostly at P2–P8 (Table 18). The P16 increment alone, interior, still correlates with Real ImageNet at 0.48–0.64 (Table 21).

(iv) Failed predictions (Table 17). They assumed P16 is written by the decoder. It is not: it is unchanged across all six decoders, which is how the input pass-through was found.

(v) The bound follows from equivariance. Agreed. The measurement adds where energy sits below s (VAR puts 77 % at P8) and what lies above s: shared, and correlated with real photographs.

W3 (mechanism evidence follows from the architecture). Point by point:

(i) The replay confirms the architecture. For AliTok alone, yes: the per-position weights are the only position-dependent parameters, and §4.2 now says so. The mechanism claim does not rest on one architecture but on a rule across twelve decoders: the operators of each (Table 20) predict which periods it can write, and the measured profiles follow — including VAR (×16), whose energy sits at P8 (77 %), which the upsampling factor alone would not predict.

(ii) The phase-averaged map in Fig. 5 is border-dominated. That is expected: averaging the weights gives each block one value, so only block borders differ. What the replay separates is weights-only (r = 0.996) from bias-only (0.26) and phase-averaged (−0.18): the pattern needs position-specific weights.

(iii) No scale selectivity in the intervention. Correct, and the scale claims do not rest on it; they rest on the operator audit and the disjoint-batch control (W2).

(iv) Profiles near the degrees-of-freedom split; no noise-floor correction. For AliTok the split is predicted, not uninformative: a linear token→16×16 projection with independent per-position weights writes a white template, whose shares follow the degree-of-freedom count (measured 0.6/5.4/21.4/72.5 % vs 1.2/4.7/18.8/75.3 %, Table 20). The template is far from noise: 20× Real's amplitude, split-half 0.997, and decoder-specific (0.996 within AliTok, ≤ 0.13 to other decoders). The noise-floor correction is now applied to all 27 cohorts (Table 22): no share moves by more than 0.049, and ×8 decoders sit far from the split (P2 at 7–46× its share).

AuthorsOpenReview ↗

Response to Reviewer K5hR (2/2): W4, W5, Q1–Q5

W4 (application results).

(i) Ranking. On C2I, Frequency-64D and the NPR probe match or pass Pixel Echo, also at an L-BFGS stationary point (Table 29). On T2I — seven commercial and open services at native resolution — Pixel Echo alone leads every baseline on the Test (90.9 vs 89.7 UnivFD, 83.1 Frequency-64D), stays ahead at stationarity (91.3 vs 76.2 NPR probe), and carries 9.5 pp inside Max (Table 31).

(ii) Detection. Low DiT recall at 1 % FPR is what the period rule predicts for a ×8 decoder (W2); attribution still recognises DiT at ≥ 91.2 % recall (Table 4). Applied jointly, the per-source heads give a 2.6 % Real FPR (Table 35).

(iii) Transformations. See Q5.

W5 (presentation). Internal labels (E29, C14), file paths and revision traces are removed, and an evidence index now opens the appendix (p. 14): one row per claim, with its main-text section, main table and supporting material.

Q1 (disjoint latent batches). Run and pre-registered; see W2 (i) in our first reply and Table 21. On the full raster the test fails because the padded ring is itself reproducible (C = 0.970, median X = 0.809), which is why the interior window was registered as primary.

Q2 (noise floor). Now applied to all 27 cohorts (Table 22, split-half cross-energy, 200 splits). It moves no published share by more than 0.049; 3 of 108 full-raster cells fall inside it, all P2. ×16 decoders sit near the degrees-of-freedom split, as a linear patch projection predicts (W3 iv); ×8 decoders sit far from it (P2 at 7–46× its share).

Q3 (stationarity, baseline concatenation). See W4 (i). The Freq ⊕ NPR concatenation is now a row of Table 31: 99.0 on C2I, 91.4 on T2I, against 98.6 and 95.0 for Max.

Q4 (Seedream; pipeline vs decoder). Seedream puts 99.8 % of its energy at P2–P8 (P16 share 0.002), the profile of a decoder of factor ≤ 8; whether part of it comes from an upscaling or watermark stage cannot be separated without a local decoder. Metadata alone attributes T2I at 43.4 %, and 20.4 % inside the 1024² block where Pixel Echo reaches 93.6 % (Table 33). The local-versus-hosted comparison you suggest is the right test; we have not collected hosted exports of the open-weight models in our set.

Q5 (shift-invariant variant). Done (Table 38). On C2I the folded-map DFT magnitude costs 0.4 pp clean (95.3 vs 95.7) and holds 95.0 under a random crop-shift, where the published feature drops to 49.6. Under bicubic resampling every variant is near chance, so the grid requirement concerns resampling, not the phase origin.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.