acceptodds
Under review as a conference paper at ICLR 2027

PRISM: Post-hoc Recognition of Image Source Models

Abstract

Image generative models (GMs) now produce photorealistic content that is virtually indistinguishable from natural images. This creates a critical need for reliable provenance to distinguish between synthetic vs authentic images and attribute generated content to its source. While diffusion models (DMs) have been the dominant paradigm for visual content generation, recent image autoregressive models (IARs) match DMs' generation quality and substantially improve the sampling speed by adapting the next-token prediction paradigm from large language models. We propose a unified post-hoc attribution method that identifies the source GM of a given image by exploiting model-specific artifacts (fingerprints) that arise in both DMs and IARs. Our key insight is to combine the fingerprints from the two main constituent parts of GMs: (1) the autoencoder that maps between pixel and latent spaces and (2) the latent generator that defines the image distribution in DMs and produces visual tokens in IARs. In the first stage, we identify the autoencoder family using reconstruction and quantization errors. In the second stage, we attribute the image to the latent generator within the identified model family using conditional log-likelihoods for DMs and token log-probabilities for IARs. Across diverse model families and versions, our two-stage framework enables accurate model attribution, providing a unified approach to the provenance of synthetic images.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
model attribution, image autoregressive model, diffusion model

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

This paper proposes PRISM, a unified post-hoc framework for attributing generated images to their source model. The attribution task is decomposed into two levels. Stage 1 identifies the model family using an autoencoder fingerprint built from the double reconstruction ratio ρRec, which the authors augment with the quantization error LQ and multiply into the family fingerprint for image autoregressive models (IARs). Stage 2 then localizes the specific version within the identified family using a latent generator fingerprint, namely an ELBO-based conditional log-likelihood for diffusion models (DMs) and token-level conditional log-probabilities for IARs. The authors evaluate 11 model families and 24 model versions, reporting 94.5% Macro-F1 for class-to-image attribution and 95.5% for text-to-image attribution, with family-level accuracies of 99.2% and 99.4%. The appendix adds experiments on robustness, open-world rejection, rejection of unseen versions, conditioning estimation, inference cost, and privacy-preserving collaborative auditing.

Critical strengths

The paper defines the task clearly, the two-level decomposition into family identification and version localization is reasonable, and the experimental coverage in the appendix goes beyond comparable post-hoc attribution work.

Critical weaknesses
  1. The binding between autoencoder type and generation paradigm is not verified, which undercuts the central claim of cross-paradigm unification. In Section 3.2 the paper states that the autoencoder is instantiated as a continuous VAE in diffusion models and as a VQVAE in IARs, and Section 3.3 uses this to determine the family fingerprint, so that DMs take LA = ρRec directly while IARs take LA = ρRec · LQ. However, the autoencoder type is not determined by the generation paradigm. Diffusion models with VQ or discrete latent autoencoders exist, and autoregressive models with continuous latent spaces also exist. More importantly, all 24 evaluated models in Table A1 fall inside this correspondence. The DM side consists entirely of continuous VAEs (SD1-VAE, SD2-VAE, SD3-VAE, SDXL-VAE, FLUX-VAE, VA-VAE, and SD-VAE), and the IAR side is built entirely on quantized or discrete tokenizers (MaskGIT-VQGAN, multi-scale VQVAE, and VQGAN-16384). The premise has therefore never been tested across the correspondence it assumes, even though applicability to both DMs and IARs is the headline claim of the paper. I also note that the paper hedges on the IAR side with "for most IARs" while making an unqualified assertion on the DM side. This asymmetry suggests that the premise has not been examined carefully.
  2. The evaluation set is constructed entirely by the authors and lacks a public benchmark, so the results cannot be compared with other work. Section 4.1 describes an evaluation set of 26,000 images generated by the authors, with 1,000 images per model across 24 models, and real images limited to 1,000 each from ImageNet and MS-COCO. No public benchmark for generated-image provenance is used anywhere in the paper, and no comparison is made against numbers reported in prior work. Because the authors re-implemented every baseline, gaps such as 94.5% versus 58.4% in Table 1 cannot be cross-checked against previously published values. In addition, the reference population for Grubbs' test, 250 images per class, and the evaluation images, the remaining 750 per class, are drawn from the same generation run and the same conditioning distribution. If the conditioning distribution shifts, for example toward prompt distributions found in the wild, the reported rejection rates are likely to be optimistic.
  3. The normality assumption behind Grubbs' test is verified only in a narrow setting. Appendix I.2 reports skewness for only nine autoencoder pairings, arranged as a 3 by 3 matrix, and all of them are measured on clean reconstruction distributions. One pairing, RAR-B against LlamaGen-L, has skewness 1.0132, which already exceeds the |skew| < 1 criterion that the authors adopt. Appendix E shows that strong augmentation substantially shifts the score distribution. At JPEG quality 60, for example, Macro-F1 drops to 81.9. Whether the Stage 1 rejection decision still holds under those conditions is not re-examined, even though the rejection mechanism depends directly on this normality assumption.
  4. Robustness and mechanism behavior are insufficiently validated in difficult settings. First, the robustness study in Appendix E covers only five conventional image transformations and is restricted to the RAR family with augmentation-finetuned encoders. It does not include compounding degradations such as repeated recompression, resizing, and screenshots applied in sequence, local edits such as cropping or inpainting, or adaptive adversarial perturbations designed by an attacker who knows the attribution method. The last of these is the most relevant threat model for a provenance method. Second, rejection of unseen versions is uneven within a family. LlamaGen-XL reaches only 71.1% at q = 1%, while FLUX derivatives sharing a single VAE are all above 99.9% under the same setting. The paper shows that the threshold is stable from q = 1% to q = 30% only for the FLUX case, offers no explanation for the degradation on the IAR side, and provides no operating-point trade-off curve. Third, Appendix D.2 shows that finetuning the encoder noticeably improves attribution, so the reported numbers depend on a one-time offline adaptation per family. The training-free claim in the main text should be scoped accordingly.
Questions for authors
  1. On the correspondence between autoencoder type and paradigm: how should LA be instantiated when a diffusion model uses a VQ or discrete autoencoder, for example a VQ-Diffusion or RQ-VAE style model, or when an autoregressive model uses a continuous latent space? Do the authors consider that this choice should be driven by the autoencoder type rather than by the generation paradigm? Could the authors add models that cross the current correspondence as a controlled comparison?
  2. On data and comparability: could the authors evaluate on a public benchmark for generated-image provenance, or at least provide a comparable comparison against numbers reported by prior work? In addition, could the Stage 1 rejection accuracy be re-verified using a reference population that is sampled independently of the query images and their conditioning distribution?
  3. On Grubbs' test: could the authors re-measure the skewness of the score distribution and the accuracy of the rejection decision under the augmentations used in Appendix E, for example JPEG quality 60 or a 0.5 times resize, and clarify whether the normality assumption still holds at those strengths? The specific effect of the RAR-B against LlamaGen-L pairing with skewness 1.0132 would also be helpful to understand.
  4. On robustness and version rejection: could the augmentation study be extended to all 24 model versions and supplemented with compounding degradations, local edits, and adaptive adversarial perturbations? For unseen-version rejection, could the authors provide a trade-off curve between the non-rejection rate and the false-rejection rate as the threshold q varies from 1% to 30%, and explain why LlamaGen-XL performs substantially worse?
  5. On privacy-preserving collaborative auditing: Appendix J specifies the full protocol and a conflict-resolution procedure, but only as a mechanism description. Could the authors provide a formal privacy guarantee or a measure of information leakage?
3 comments
AuthorsOpenReview ↗

Authors' Response (1/3)

Thank you for your constructive feedback! We address your main concerns as follows.

Architecture and training (W1/Q1). The quantizer determines whether a quantization error is available. For example, a VQ diffusion model could use the quantized Stage 1 fingerprint, and a continuous-latent AR model could use the reconstruction ratio with a continuous generator score in Stage 2. Table A1 covers continuous-autoencoder DMs and discrete-tokenizer IARs. The source generator stays fixed, and the scoring encoder can be adapted offline.

Benchmark and calibration (W2/Q2). As far as we understand, there are no comprehensive benchmarks that cover all the models we evaluate. Therefore, we construct this benchmark for the provenance task, which is also one of our contributions to the community. Every method faces the same checkpoints and held-out images, making Table 1 a controlled version-attribution comparison.

Gaussian assumption (W3/Q3). PRISM’s central contribution is the selection and combination of complementary autoencoder and generator signals for fine-grained attribution. We chose Grubbs test as an instantiation of the statistical test, but the specific test is replaceable. In particular, a KDE-based autoencoder compatibility test estimates the reference-score distribution without requiring Gaussianity. We evaluated a KDE-based variant of PRISM. As shown below, it yields attribution recall close to the original Grubbs-based implementation across the 13 evaluated generators, while maintaining real-image rejection rates of 97.6% on ImageNet and 98.9% on MSCOCO. This supports the use of PRISM’s attribution signals with an alternative autoencoder compatibility test that does not require Gaussian reference distributions.

Table R1. Grubbs versus KDE on the class-to-image pool. The results for Grubbs test correspond to Figure 3a in the paper.

SourceMetricGrubbs (%)KDE (%)
RAR-BAttribution recall99.5 ± 0.399.3 ± 0.3
RAR-LAttribution recall96.9 ± 0.697.2 ± 0.6
RAR-XLAttribution recall92.1 ± 1.091.7 ± 1.0
RAR-XXLAttribution recall90.4 ± 1.190.5 ± 1.1
VAR-16Attribution recall98.3 ± 0.596.8 ± 0.6
VAR-20Attribution recall87.3 ± 1.285.7 ± 1.3
VAR-24Attribution recall86.5 ± 1.285.9 ± 1.3
VAR-30Attribution recall91.9 ± 1.091.6 ± 1.0
LlamaGen-BAttribution recall93.9 ± 0.993.5 ± 0.9
LlamaGen-LAttribution recall95.1 ± 0.893.9 ± 0.9
LlamaGen-XLAttribution recall94.5 ± 0.891.9 ± 1.0
LightningDiTAttribution recall99.2 ± 0.398.1 ± 0.5
REPAAttribution recall99.1 ± 0.497.9 ± 0.5
ImageNetCorrect rejection97.5 ± 0.697.6 ± 0.6
MSCOCOCorrect rejection98.8 ± 0.498.9 ± 0.4
AuthorsOpenReview ↗

Authors' Response (2/3)

Robustness (W4/Q4). In the following extensive table, we evaluate the robustness of our approach, across 21 settings spanning five perturbation types. With the augmentation-finetuned encoder, PRISM retains per-version F1 of 82.2–86.6 at JPEG quality 60 and 83.7–88.3 after 0.5× resizing. These results demonstrate that fine-grained attribution can remain effective after substantial post-processing without relying on a watermark.

We would like to further note that robustness against post-processing can be a very challenging aspect for provenance, which is hard even for methods that deliberately embed watermarks. Our approach faces the additional constraint of passive attribution: it must use existing model traces without modifying generation to protect them. The main focus of our paper is building a fine-grained, unified attribution method rather than making the method robust to all possible transformations. We will update more results for robustness evaluation in the camera-ready version.

Table R2. Robustness to post-processing, by augmentation type and strength. We report F1 per model version with the augmentation-finetuned encoder.

AugmentationStrengthRAR-BRAR-LRAR-XLRAR-XXLMacro-F1
None—94.595.095.594.796.2
JPEGq=9093.592.194.392.693.4
JPEGq=8091.690.992.790.290.4
JPEGq=7089.085.290.388.186.8
JPEGq=6086.582.286.682.781.9
Resize×0.994.594.796.494.696.2
Resize×0.893.593.295.293.395.2
Resize×0.793.694.095.692.994.9
Resize×0.691.991.094.590.592.7
Resize×0.586.583.788.384.684.4
Gaussian blurr=0.594.894.996.094.796.4
Gaussian blurr=1.091.692.295.191.893.0
Gaussian noisestd=194.193.995.393.695.3
Gaussian noisestd=293.993.194.692.994.2
Gaussian noisestd=392.092.093.391.192.2
Gaussian noisestd=492.090.792.889.790.5
Gaussian noisestd=590.187.990.188.487.4
Saturation×1.194.594.195.394.495.5
Saturation×1.293.193.794.493.493.6
Saturation×1.391.591.293.192.291.1
Saturation×1.489.288.490.589.587.3
Saturation×1.585.785.187.086.082.9

Unseen versions (W4/Q4). PRISM checks how decisively the top-scoring version beats the runner-up. It returns “family known, version unknown” when this margin falls below the q-quantile of margins on the winner’s own calibration images. Increasing q raises the threshold, rejecting more unknown images at the cost of more false rejection.

The LlamaGen-XL result reflects the remaining candidates. With XL held out, every XL image scores higher under L than B, and 28.9% retain a margin above the q = 1% threshold. FLUX leaves six siblings, whose top two scores are typically nearly tied, making rejection easier.

Table R3 shows the trade-off. At q = 5%, XL rejection rises from 71.1% to 89.3%. Across LlamaGen leave-one-out settings, average rejection rises from 89.6% to 96.4%, while false rejection increases from 1.9% to 6.7%.

Table R3. Unknown-version rejection and false rejection as q varies. Each version is held out in turn and tested on its 750 query images. All rates are percentages, rounded to one decimal place.

Metric (%)q=1%q=2%q=5%q=10%q=20%q=30%
FLUX: unknown-version rejection100.0100.0100.0100.0100.0100.0
FLUX: false rejection1.32.86.812.724.434.8
LlamaGen: unknown-version rejection89.691.196.497.999.399.6
LlamaGen: false rejection1.93.26.711.821.531.2
LlamaGen-XL held out: rejection71.174.189.393.798.098.8
Family rows are averages over leave-one-version-out settings. False rejection is measured among correctly attributed enrolled images within those settings.

If our response addresses your concerns, we would greatly appreciate it if you could consider raising the score. Thank you!

AuthorsOpenReview ↗

Authors' Response (3/3)

Privacy guarantee (Q5). We formalize the information disclosed by the collaborative protocol in Appendix J.

Guarantee. Let P denote public protocol parameters and θ the owners’ private model weights, reference data, and calibration statistics. For a query x, define the disclosed output L=(r_1,...,r_m,b_1,...,b_t), where each r_i is an owner’s categorical response and each b_j is a disclosed conflict-comparison outcome. Assume an honest-but-curious coordinator that does not collude with owners, sequentially executed comparisons with simulation security for the coordinator’s visible messages, and communication schedules and message lengths determined by P, x, and L.

Let View_C(P,x,θ) denote the coordinator’s view in a real execution, including its inputs, random coins, and received protocol messages. There exists a probabilistic polynomial-time algorithm Sim that takes only P, x, and L, has no access to θ, and outputs a simulated coordinator view such that

View_C(P,x,θ) ≈c Sim(P,x,L).

Here, ≈c denotes computational indistinguishability. Thus, the coordinator learns no additional computationally accessible information about the private state beyond the disclosed outputs.

Proof. We construct Sim(P,x,L) as follows. The simulator samples the coordinator’s random coins according to the protocol and reproduces the query-distribution messages. It then inserts the categorical replies r_1,...,r_m supplied in L. These messages are reproduced exactly. Autoencoder evaluation, generator scoring, and calibration are local computations and contribute no further messages.

For each conflict comparison, the simulator uses the comparison protocol’s simulator with its public parameters and disclosed outcome b_j. The communication schedule and message lengths are reproducible from the stated inputs. Consider hybrid transcripts that replace real comparison executions with these simulated executions one at a time. Any efficient distinguisher between two adjacent hybrids would distinguish a real comparison execution from its simulation, contradicting the assumed comparison security. Since the protocol uses polynomially many sequential comparisons, the sum of these distinguishing advantages is negligible. The simulator outputs the assembled coordinator view. The first hybrid is the real coordinator view and the last is the distribution of Sim(P,x,L). This proves the claimed computational indistinguishability.

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

This paper proposes PRISM, a unified post-hoc source attribution framework for both Diffusion Models and Image Autoregressive Models. It first identifies the model family using autoencoder fingerprints, and then determines the specific model version using conditional probability signals from the latent generator; experiments on 24 generative models and real images show that the proposed two-stage approach clearly outperforms existing post-hoc provenance methods in fine-grained model attribution and unknown-source rejection.

Critical strengths
  1. The overall framework is clearly designed and has practical value. The authors decompose source attribution into two stages, namely autoencoder family identification and latent generator attribution, and apply this framework uniformly to both diffusion and autoregressive models. The overall idea is intuitive and easy to interpret.
  2. Strong empirical performance: The paper evaluates the method on 24 different generative models, covering multiple model families, multiple model versions, and real images. In the main experiments, the proposed method clearly outperforms existing post-hoc attribution baselines.
Critical weaknesses
  1. Novelty: The ECCV 2026 MGI paper(MGI: Member vs Generated Inference) already adopts a highly similar staged attribution strategy based on “autoencoder fingerprint + latent-generator likelihood.” The authors should further clarify the substantive methodological novelty of PRISM over this work, rather than emphasizing only differences in task formulation and experimental scale.
  2. Limited robustness evaluation: The current robustness experiments are conducted only on RAR and use an augmentation-finetuned encoder. I suggest evaluating the robustness of the method across multiple diffusion and IAR families.
  3. The current open-world evaluation is not fully convincing: The experiments mainly focus on rejecting models from completely different families, while a more realistic and challenging setting is an unseen fine-tune or new model version that shares the same autoencoder with an enrolled model. I suggest adding experiments under this setting.
Questions for authors

N/A

2 comments
AuthorsOpenReview ↗

Authors' Response (1/2)

Thank you for your constructive feedback! We would like to address your concerns as follows.

W1. Novelty. The main methodological differences between MGI and PRISM are:

(1) Unknown-version rejection within a known family. PRISM adds a margin-based rejection rule calibrated on the winning candidate’s own reference images. This enables “family known, version unknown” decisions when an unregistered fine-tune shares an enrolled model’s autoencoder, extending attribution beyond MGI’s comparison between given candidate sources.

(2) Key change in model version attribution. MGI fits source densities for a candidate pair. PRISM first identifies candidate model families from autoencoder scores, then ranks candidate versions directly by their conditional generator scores. This replaces pairwise source-density fitting with a direct ranking rule for fine-grained version attribution. Figure A5 shows why this choice matters: on the autoregressive models, direct conditional scoring outperforms both token weighting and conditional–unconditional differencing.

(3) Unified attribution across families and versions. PRISM extends MGI’s parent–derivative comparison to source attribution across diffusion and autoregressive models, including unregistered models, model versions, and fine-tuned models. The framework combines model family identification, version ranking, and rejection at both family and version levels. Model family identification also provides a computational benefit. With 11 candidates, PRISM takes 0.387 seconds per image versus 1.367 for PRADA (Table R1 below).

(4) Comprehensive benchmark. We additionally contribute a comprehensive image-provenance benchmark spanning 24 diffusion and autoregressive model versions in class-to-image and text-to-image settings, with 26,000 images: 1,000 per model and 1,000 each from ImageNet and MS-COCO.

Table R1. Average inference time (seconds per image) comparison with PRADA.

# ModelsPRADAPRISM
4 (RARs only)0.3500.159
8 (RARs + VARs)1.0190.291
11 (RARs + VARs + LlamaGen)1.3670.387

W2. Robustness. In the following extensive table, we evaluate the robustness of our approach, across 21 settings spanning five perturbation types. With the augmentation-finetuned encoder, PRISM retains per-version F1 of 82.2–86.6 at JPEG quality 60 and 83.7–88.3 after 0.5× resizing. These results demonstrate that fine-grained attribution can remain effective after substantial post-processing without relying on a watermark.

We would like to further note that robustness against post-processing can be a very challenging aspect for provenance, which is hard even for methods that deliberately embed watermarks. Our approach faces the additional constraint of passive attribution: it must use existing model traces without modifying generation to protect them. The main focus of our paper is building a fine-grained, unified attribution method rather than making the method robust to all possible transformations. We will update more results for robustness evaluation in the camera-ready version.

Table R2. Robustness to post-processing, by augmentation type and strength. We report F1 per model version with the augmentation-finetuned encoder.

AugmentationStrengthRAR-BRAR-LRAR-XLRAR-XXLMacro-F1
None—94.595.095.594.796.2
JPEGq=9093.592.194.392.693.4
JPEGq=8091.690.992.790.290.4
JPEGq=7089.085.290.388.186.8
JPEGq=6086.582.286.682.781.9
Resize×0.994.594.796.494.696.2
Resize×0.893.593.295.293.395.2
Resize×0.793.694.095.692.994.9
Resize×0.691.991.094.590.592.7
Resize×0.586.583.788.384.684.4
Gaussian blurr=0.594.894.996.094.796.4
Gaussian blurr=1.091.692.295.191.893.0
Gaussian noisestd=194.193.995.393.695.3
Gaussian noisestd=293.993.194.692.994.2
Gaussian noisestd=392.092.093.391.192.2
Gaussian noisestd=492.090.792.889.790.5
Gaussian noisestd=590.187.990.188.487.4
Saturation×1.194.594.195.394.495.5
Saturation×1.293.193.794.493.493.6
Saturation×1.391.591.293.192.291.1
Saturation×1.489.288.490.589.587.3
Saturation×1.585.785.187.086.082.9
AuthorsOpenReview ↗

Authors' Response (2/2)

W3. Unseen fine-tunes and new versions sharing an autoencoder. We evaluate both requested cases by removing one model from the candidate set and testing its 750 held-out images, while keeping sibling models enrolled. These images therefore come from a known autoencoder family but an unknown generator. PRISM identifies this case using generator-score margins: when the top-two margin is below the winner’s calibrated 1st percentile (q = 1%), it returns “family known, version unknown.”

(1) Unseen fine-tunes. Our seven FLUX models share an identical VAE and cover LoRA adaptation, full fine-tuning, and distillation. PRISM rejects 99.9–100% of images from each held-out model, with 1.3% average false rejection of correctly attributed enrolled images.

Table R3. Held-out FLUX models sharing an identical VAE.

Held-out versionNo rejection: mostly labelled modelNo rejection: labelled rate (%)With rejection: true rejection rate (%)
FLUX.1-DevMidJourney-LoRA66.0100.0
AWPortrait-LoRAAWPortrait-Full99.2100.0
AWPortrait-FullAWPortrait-LoRA98.999.9
KreaDevAWPortrait-Full83.2100.0
HyperSD-LoRAMidJourney-LoRA37.9100.0
MidJourney-LoRAAWPortrait-Full68.1100.0
schnellAWPortrait-Full66.0100.0

(2) Unseen model versions. We apply the same protocol to LlamaGen-B/L/XL, which share the same tokenizer. Rejection rates are shown below. Average false rejection of correctly attributed enrolled images is 1.9%. Thus, the evaluation explicitly tests new versions within an enrolled family as well as unseen fine-tunes.

Table R4. Held-out LlamaGen versions sharing the same tokenizer.

Held-out versionNo rejection: mostly labelled modelNo rejection: labelled rate (%)With rejection: true rejection rate (%)
LlamaGen-BLlamaGen-L97.998.3
LlamaGen-LLlamaGen-XL98.899.3
LlamaGen-XLLlamaGen-L100.071.1

If our response addresses your concerns, we would greatly appreciate it if you could consider raising the score. Thank you!

Reviewer_2LYGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PRISM, a post-hoc method that attributes a generated image to its source model across both latent diffusion models (DMs) and image autoregressive models (IARs). Stage 1 identifies the model family via an autoencoder fingerprint (a double-reconstruction ratio, multiplied by the VQ quantization error for IARs), accepting a family when the score passes a Grubbs test against 250 reference images. Stage 2 then selects the version within that family with the highest conditional likelihood (token log-probability for IARs, denoising ELBO for DMs), using conditions estimated by EVA02 or BLIP-2. Across 24 models and two real-image sets, the paper reports 94.5 (class-to-image) and 95.5 (text-to-image) macro-F1, against 58.4 and 7.9 for PRADA.

Critical strengths

S1. Clear decomposition. Separating family attribution (autoencoder) from version attribution (latent generator) is well motivated by Fig. 2, and Fig. 1 communicates the design clearly.

S2. One framework for both paradigms. The method covers DMs and IARs, including seven FLUX derivatives sharing one VAE (Fig. A8), where Stage 2 carries the full decision.

S3. A useful simplification. The finding that a plain conditional likelihood outperforms calibrated membership-inference scores (Tab. 1, App. D.3) is informative, and the hierarchical design is 2.2-3.5x faster than PRADA (Tab. A5).

Critical weaknesses

W1. Robustness is evaluated on a single family. Tab. A3 covers only the four RAR versions, relies on an "augmentation-finetuned encoder" that is not described anywhere, and reports no real-image rejection under perturbation. The Macro-F1 column also averages over classes that are not shown: it is 96.2 in the "None" row, above all four per-version scores (94.5-95.5), but 81.9 at JPEG q=60, below all four (82.2-86.6). Put differently, whatever is hidden degrades fastest. The same analysis is needed for at least one DM and one text-to-image family, including rejection rates and realistic processing chains (e.g., resize followed by JPEG).

W2. No threat model. Sec. 3.1 motivates PRISM with incident response and suspending versions from a registry, where a false attribution is costly. However, Stage 1 accepts any image that resembles a decoder output of A_i, and a single encode-decode pass through a public autoencoder produces exactly that (cf. SD "laundering" of real images, Mandelli et al., arXiv 2407.10736). A third party could hence frame an enrolled model at negligible cost. The paper should either evaluate laundering and standard evasion attacks (Carlini & Farid, CVPRW 2020) or state explicitly that it assumes a non-adversarial setting.

W3. The Stage 1 test is framed in reverse, and its assumptions are not verified. Eq. 7-8 set H0 as "x was not generated by A_i", but Grubbs' test detects whether x is an outlier relative to A_i's own outputs; failing to detect an outlier is not evidence of membership. Hence, α bounds how often A_i's own images are rejected, not how often other images are falsely accepted, which is the error that matters here. With N=250 and α=0.01, the rule reduces to a fixed threshold of z < 3.89, so an empirical-quantile threshold would give the same decision without the Gaussian assumption (which Tab. A7 checks only via skewness, on 3 autoencoders, with one exceeding the paper's own threshold at 1.0132). Relatedly, Alg. 1 takes an argmax across all passing families, comparing an ELBO against summed token log-probabilities, and the class- and text-to-image pools are never combined, so possible collisions (e.g., REPA's SD-VAE vs. SD1-VAE, Tab. A1) remain untested.

W4. Possible shortcuts in the data. Per Sec. 3.5 and 4.4, versions sharing an autoencoder should be inseparable at Stage 1, as observed for SD-1.4/1.5 (autoencoder-only F1 of 66.0/0.2). Yet LlamaGen-XL Stage 1/2 reach 93.8/99.1 (Tab. A2) despite sharing the VQGAN-16384 tokenizer. Notably, LlamaGen generates Stage 1 at 256x256 and Stage 2 at 512x512, suggesting the separation stems from resolution rather than the autoencoder. Similarly, the real images are JPEG-compressed while the storage format of generated images is not stated; if the latter are PNGs, rejection may partly reflect compression, a known dataset bias (Grommelt et al., arXiv 2403.17608).

W5. Stage 2 depends on generation and conditioning details. App. F describes the effect of EVA02-estimated labels as minimal, but comparing Fig. A6a and A6b, RAR-XXL drops from 99.2% to 90.4% and VAR-d30 from 99.5% to 91.9%. Calibration and test images also share generation settings and prompt sources, while the likelihood is plausibly sensitive to guidance scale, sampler, and IAR temperature/top-k; none of these are varied. Lastly, Eq. 4 assumes ε-prediction, and it is unclear how it was computed for rectified-flow models (SD-3.5, FLUX).

W6. Missing related work and overstated novelty. Related work covers only the most recent DM-autoencoder and IAR methods. Missing are: a 2025 attribution method of the same name (PRISM, Ricco et al., arXiv 2509.15270); GAN fingerprinting (Yu et al., ICCV 2019); open-set attribution (Girish et al., ICCV 2021; POSE, CVPR 2023), against which App. G should be positioned; single-model attribution via inversion (FLIPAD, ICML 2024), the closest prior to Stage 1; attribution across fine-tunes of one base model (Wißmann et al., ACM MAD 2024); DE-FAKE (CCS 2023), which also recovers prompts with BLIP; and reconstruction- or likelihood-based detectors (DIRE, ICCV 2023; ZED, ECCV 2024). Accordingly, the claims of "the first unified attribution method" (Sec. 1) and of no prior version-level score for DMs (Tab. A11) should be narrowed, as the conditional denoising loss is already used as a decision score by Li et al. (2023a) and Zhai et al. (2024), both cited.

W7. Figs. 3 and 4 are illegible. Each matrix has 14-18 classes per axis with two numbers per cell (the same applies to Figs. A1, A2, and A4-A8). The ± values carry no information, as they equal the binomial standard error √(p(1-p)/750) (e.g., 87.3 ± 1.2, 50.1 ± 1.8). I suggest dropping them, printing only off-diagonal entries ≥1%, and replacing the main-text matrices with a per-version accuracy table. By contrast, Fig. 1 is clear and does real explanatory work.

W8. Reference formatting. Capitalization is lost throughout the bibliography (e.g., Maskgit, Prada, Aeroblade, Blip-2, Imagenet, "Ai models collapse", "Flux. 1 kontext"); protecting names with braces in BibTeX fixes this. Additionally, Alemohammad et al. is listed under the Twelfth ICLR in 2023 (held in 2024), and FLUX.1-dev is cited via the FLUX.1 Kontext paper, which describes an editing model.

W9. Method name. I like the name, but "PRISM" is already associated with the PRISM Alignment Dataset (Kirk et al., NeurIPS 2024 D&B best paper) and, more importantly, with the attribution method in W6. This is only a suggestion and does not affect my score.

Questions for authors

Q1. (W3) What are the results on a single pool of all 24 models? How often does more than one family pass Stage 1, and how are cross-family L_G scores compared in that case?

Q2. (W4) Why does the autoencoder-only variant separate LlamaGen-XL Stage 1 and 2? What are the results with all images at a common resolution, and in which format were generated images stored?

Q3. (W2) What is the false-attribution rate for real images passed once through an enrolled autoencoder, and under each perturbation in Tab. A3?

Q4. (W1) Which augmentations were used for the "augmentation-finetuned encoder", and do they overlap with the test perturbations?

Q5. (W5) How was Eq. 4 computed for rectified-flow and guidance-distilled models, and with how many timesteps and noise draws per image?

Q6. Sec. 3.1 motivates separating an owner's outputs from third-party fine-tunes, yet without App. M's margin test, 37.9-100% of a held-out version's images are assigned to an enrolled sibling (Tab. A10). Should the margin test be part of the main method?

4 comments
AuthorsOpenReview ↗

Authors' Response (1/4)

Thank you for your constructive and extensive feedback! We would like to address each point as follows.

W1. Robustness. In the following extensive table, we evaluate the robustness of our approach, across 21 settings spanning five perturbation types. With the augmentation-finetuned encoder, PRISM retains per-version F1 of 82.2–86.6 at JPEG quality 60 and 83.7–88.3 after 0.5× resizing. These results demonstrate that fine-grained attribution can remain effective after substantial post-processing without relying on a watermark.

We would like to further note that robustness against post-processing can be a very challenging aspect for provenance, which is hard even for methods that deliberately embed watermarks. Our approach faces the additional constraint of passive attribution: it must use existing model traces without modifying generation to protect them. The main focus of our paper is building a fine-grained, unified attribution method rather than making the method robust to all possible transformations. We will update more results for robustness evaluation in the camera-ready version.

Table R1. Robustness to post-processing, by augmentation type and strength. We report F1 per model version with the augmentation-finetuned encoder. Macro-F1 averages four RAR-version F1s and two natural-dataset F1s. Both natural datasets use the Outlier prediction. When computing either natural-dataset F1, images from the other natural dataset are excluded from its false-positive count.

AugmentationStrengthRAR-BRAR-LRAR-XLRAR-XXLMacro-F1
None—94.595.095.594.796.2
JPEGq=9093.592.194.392.693.4
JPEGq=8091.690.992.790.290.4
JPEGq=7089.085.290.388.186.8
JPEGq=6086.582.286.682.781.9
Resize×0.994.594.796.494.696.2
Resize×0.893.593.295.293.395.2
Resize×0.793.694.095.692.994.9
Resize×0.691.991.094.590.592.7
Resize×0.586.583.788.384.684.4
Gaussian blurr=0.594.894.996.094.796.4
Gaussian blurr=1.091.692.295.191.893.0
Gaussian noisestd=194.193.995.393.695.3
Gaussian noisestd=293.993.194.692.994.2
Gaussian noisestd=392.092.093.391.192.2
Gaussian noisestd=492.090.792.889.790.5
Gaussian noisestd=590.187.990.188.487.4
Saturation×1.194.594.195.394.495.5
Saturation×1.293.193.794.493.493.6
Saturation×1.391.591.293.192.291.1
Saturation×1.489.288.490.589.587.3
Saturation×1.585.785.187.086.082.9

The two natural-dataset F1s explain the Macro-F1 values: ImageNet/MSCOCO score 98.6/98.7 on clean images and 76.7/76.9 at JPEG quality 60. Their F1 decreases mainly because more generated images are rejected as Outlier, reducing Outlier precision. Across every condition in this RAR-only table, real-image false attribution remains at most 0.67% for ImageNet and 0.40% for MSCOCO.

W2. Threat model and closed-model deployment. PRISM supports owner-side auditing of closed models: the owner computes both stages locally, while an external auditor receives only the attribution decision (Sec. 3.1, App. J). If the autoencoder is also private and no reconstruction interface is exposed, the proposed direct encode–decode laundering operation is unavailable to the third party. This access restriction does not apply to a closed generator using a public VAE. Our evaluated scope is mostly non-adversarial source attribution, and we will update the scope in the paper to clarify this. Stage 1 tests autoencoder compatibility. Final version attribution additionally uses generator scores and, for an open registry, margin-based rejection.

AuthorsOpenReview ↗

Authors' Response (2/4)

W3. Statistical interpretation, autoencoder compatibility tests, and candidate collisions.

(1) Interpretation. Stage 1 tests autoencoder compatibility against the reference distribution. Candidate families that pass proceed to Stage 2. We will state this directly in Eqs. 7–9. The Grubbs significance level concerns outlier detection under the reference model. False attribution to an unrelated source is measured empirically by the full pipeline, rather than bounded by α.

(2) Gaussianity. PRISM’s central contribution is the selection and combination of complementary autoencoder and generator signals for fine-grained attribution. Grubbs’ test is one possible autoencoder compatibility test. When Gaussianity is unsuitable, it can be replaced by a KDE-based autoencoder compatibility test, which estimates the reference-score distribution without assuming a Gaussian form. We evaluated this replacement. Table R2 shows attribution recall close to the Grubbs-based implementation across all 13 generators, with real-image rejection rates of 97.6% on ImageNet and 98.9% on MSCOCO. These results demonstrate that PRISM’s attribution signals remain effective with an autoencoder compatibility test that does not require Gaussianity.

Table R2. Grubbs versus KDE on the class-to-image pool. The results for Grubbs test correspond to Figure 3a in the paper.

SourceMetricGrubbs (%)KDE (%)
RAR-BAttribution recall99.5 ± 0.399.3 ± 0.3
RAR-LAttribution recall96.9 ± 0.697.2 ± 0.6
RAR-XLAttribution recall92.1 ± 1.091.7 ± 1.0
RAR-XXLAttribution recall90.4 ± 1.190.5 ± 1.1
VAR-16Attribution recall98.3 ± 0.596.8 ± 0.6
VAR-20Attribution recall87.3 ± 1.285.7 ± 1.3
VAR-24Attribution recall86.5 ± 1.285.9 ± 1.3
VAR-30Attribution recall91.9 ± 1.091.6 ± 1.0
LlamaGen-BAttribution recall93.9 ± 0.993.5 ± 0.9
LlamaGen-LAttribution recall95.1 ± 0.893.9 ± 0.9
LlamaGen-XLAttribution recall94.5 ± 0.891.9 ± 1.0
LightningDiTAttribution recall99.2 ± 0.398.1 ± 0.5
REPAAttribution recall99.1 ± 0.497.9 ± 0.5
ImageNetCorrect rejection97.5 ± 0.697.6 ± 0.6
MSCOCOCorrect rejection98.8 ± 0.498.9 ± 0.4

(3) Multiple passing families. Multiple families pass the autoencoder compatibility test for 2.53% of class-to-image queries and 11.12% of text-to-image queries. Stage 2 selects the highest generator score among candidates that pass Stage 1.

W4. LlamaGen tokenizer configurations. For LlamaGen, we followed the setup in their paper: the two text-to-image configurations share decoder and codebook weights, but use different encoders in the evaluated setup: the 256-pixel configuration uses the original encoder, whereas the 512-pixel configuration uses a finetuned encoder. For identical-autoencoder version discrimination, the seven FLUX derivatives provide the direct test: they share the same VAE, while generator-level scoring distinguishes their versions and the margin test rejects unseen derivatives at 99.9–100% (Table R3). Storage and preprocessing details are given in Q2.

Table R3. Unseen FLUX derivatives, seven enrolled versions with an identical VAE.

Held-out versionNo rejection: mostly labelled modelNo rejection: labelled rate (%)With rejection: true rejection rate (%)
FLUX.1-DevMidJourney-LoRA66.0100.0
AWPortrait-LoRAAWPortrait-Full99.2100.0
AWPortrait-FullAWPortrait-LoRA98.999.9
KreaDevAWPortrait-Full83.2100.0
HyperSD-LoRAMidJourney-LoRA37.9100.0
MidJourney-LoRAAWPortrait-Full68.1100.0
schnellAWPortrait-Full66.0100.0
AuthorsOpenReview ↗

Authors' Response (3/4)

W5. Conditioning and generation settings. PRISM’s end-to-end results use estimated conditions, matching the intended setting where the original label or prompt is unavailable. Figure A6 isolates the effect of this estimation. Calibration and evaluation use disjoint subsets of the same generated dataset (250 reference and 750 test images per version), with the same generation settings. We will describe its version-dependent impact explicitly and document the generation and scoring settings alongside the results.

W6. Methodological distinctions from prior work. PRISM combines autoencoder-based family identification, direct conditional generator-score ranking, and winner-calibrated rejection of unseen sibling versions across diffusion and autoregressive models. The distinctions from each cited work are:

(1) Ricco et al. (PRISM, 2025): radial Fourier amplitude/phase signatures with linear discriminant analysis. Our method instead interrogates the candidate autoencoder and generator, separating family evidence from version evidence.

(2) Yu et al. (2019): learned GAN fingerprints support fine-grained attribution. PRISM uses candidate-model reconstruction and conditional scores for diffusion/IAR attribution, including versions sharing a decoder.

(3) Girish et al. (2021): iterative classifier training, unknown-source clustering, and discovery. PRISM tests autoencoder compatibility for enrolled models directly and makes family-level and version-level rejection decisions for each query.

(4) POSE (2023): learns an open-set classifier using progressively simulated unknown-model traces. PRISM calibrates on enrolled-model reference scores and rejects unseen siblings through winner-specific margins. Appendix G will be positioned within this open-set literature.

(5) FLIPAD (2024): final-layer inversion followed by anomaly detection for single-model attribution. PRISM uses an autoencoder compatibility test followed by conditional generator scoring to resolve versions that share that autoencoder.

(6) Wißmann et al. (2024): a trained classifier on spectral features distinguishes diffusion fine-tunes. PRISM directly ranks candidate generators and additionally returns “family known, version unknown” for unregistered derivatives.

(7) DE-FAKE (2023): image/text features, including BLIP captions, feed an attribution classifier. In PRISM, the estimated caption conditions each candidate generator’s own likelihood evaluation.

(8) DIRE (2023): diffusion reconstruction error supports real/generated detection. PRISM combines autoencoder-based family identification with source-specific generator scores for family and version attribution.

(9) ZED (2024): detects synthetic images through entropy statistics from a real-image probability model. PRISM scores candidate source generators to identify their family and version.

(10) Li et al. (2023a): compares conditional diffusion scores across semantic labels within a model. PRISM compares candidate versions for source attribution after model family identification.

(11) Zhai et al. (2024): conditional-likelihood discrepancy measures training-data membership. PRISM uses the direct conditional score to rank source versions. Our ablations compare this choice with discrepancy-based scoring.

We will make these distinctions explicit in the related work section, crediting the established scoring ingredients and specifying our contribution as their attribution-specific combination with within-family unknown-version rejection.

W7. Figure readability. The ± entries are binomial standard errors over 750 images per row. We will remove them from the dense matrices, show off-diagonal values only above 1%, and provide a compact per-version summary in the main text.

W8. References. We will correct model and dataset capitalization in BibTeX, correct the ICLR year, and cite FLUX.1-dev’s model documentation directly.

W9. Method name. We will cite the existing attribution method named PRISM and use another name. Thanks for the suggestion.

Q1. Multiple families passing Stage 1. More than one model family passes Stage 1 for 2.53% of test images in the class-to-image setting and 11.12% in the text-to-image setting. These are intermediate acceptance rates, not final misclassification rates. After Stage 2, final version-level attribution achieves Macro-F1 scores of 94.5% and 95.5% in the class-to-image and text-to-image settings, respectively.

Q2. Resolution and storage. The LlamaGen encoder distinction is explained in W4. Preprocessing uses bilinear short-side resizing with antialiasing, center cropping, and scaling to [-1, 1], applied identically to real and generated images within each extraction job. Generated images are stored as lossless PNGs. Real images are decoded from JPEG and resaved as PNGs, which is a standard approach for processing real images.

AuthorsOpenReview ↗

Authors' Response (4/4)

Q3. False attribution under perturbation. We directly evaluate whether post-processing causes real images to be falsely attributed to enrolled generators, using the robustness setting in Appendix E. With calibration fixed on clean images, false-positive rates remain at or below 0.67% on ImageNet and 0.40% on MSCOCO across the evaluated perturbations (Table R4). These results are obtained without version-level rejection, demonstrating that the underlying two-stage attribution procedure already maintains low false attribution under these transformations.

Table R4. Real-image false attribution under perturbation.

ConditionImageNet false attribution (%)MSCOCO false attribution (%)
Clean0.270.00
JPEG quality 600.400.00
Resize ×0.50.670.13
Gaussian blur r=1.00.530.40
Gaussian noise std=50.400.13
Saturation ×1.50.130.27

Q4. Encoder augmentation. Encoder adaptation is performed offline on generated images while keeping the source generator fixed. The augmentation-training implementation includes JPEG compression, blur, Gaussian noise, brightness, saturation, resizing, and contrast. It applies augmentation with probability 0.5 per batch, sampling one transform and one strength uniformly. Its training ranges overlap the evaluated JPEG, resizing, saturation, noise, and blur conditions.

Q5. Flow-model scoring. Our FLUX scoring implementation evaluates conditional velocity-prediction error: for z_t = (1−t)z₀ + tε, the target is ε−z₀. It averages prediction error over a uniform timestep grid, with one noise draw per timestep. Guidance-distilled FLUX-dev uses a guidance embedding of 3.5. The conditional and unconditional passes are evaluated separately, without a classifier-free-guidance combination. We will distinguish this flow-specific score from the ε-prediction expression in Eq. 4.

Q6. Should the margin test be part of the main method? Yes. For open-registry attribution, it belongs in the main decision rule. After model family identification and version ranking, PRISM returns “family known, version unknown” if the top-two score margin is below the winner’s q-quantile of calibration margins. The forced assignments in Tables R3 and R5 are the ablation without this rejection step. With it, q = 1% rejects 99.9–100% of unseen FLUX derivatives and 98.3%/99.3%/71.1% of held-out LlamaGen-B/L/XL images, at average false-rejection rates of 1.3%/1.9% for FLUX/LlamaGen. We will include this step explicitly in the main algorithm and report its q-dependent trade-off.

Table R5. Unseen LlamaGen versions, three enrolled versions.

Held-out versionNo rejection: mostly labelled modelNo rejection: labelled rate (%)With rejection: true rejection rate (%)
LlamaGen-BLlamaGen-L97.998.3
LlamaGen-LLlamaGen-XL98.899.3
LlamaGen-XLLlamaGen-L100.071.1

If our response addresses your concerns, we would greatly appreciate it if you could consider raising the score. Thank you!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.