acceptodds
Under review as a conference paper at ICLR 2027

DUO: Dual Readouts of Diffusion Models for OOD Detection

Abstract

Diffusion models are powerful generative methods for learning and sampling from complex data distributions. Safely deploying these models requires detecting inputs that fall outside their training distribution. Because diffusion models recover clean samples from noisy inputs, reconstruction error is often used for out-of-distribution (OOD) detection. This approach assumes that in-distribution (ID) inputs are reconstructed more accurately than OOD inputs. However, OOD inputs may be reconstructed equally well or even better, making this score unreliable. We therefore propose the leading sensitivity spectrum of intermediate denoising features as a complementary OOD signal and demonstrate that it remains informative when reconstruction error is misleading. We introduce DUal-readout of diffusion models for OOD detection (DUO), a fully unsupervised, training-free OOD detector that combines reconstruction error and feature sensitivity. DUO improves average AUROC by approximately 6% points over the strongest baselines while substantially reducing inference cost, and achieves SOTA on near-OOD and high-resolution benchmarks.

Primary area:
generative models
Keywords:
OOD detection, diffusion models, uncertainty estimation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_DNMxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper tackles with unsupervised OOD detection. It uses two scores to provide final OOD score. Given a test image, with some noise inserted, it is measured how well diffusion model is able to reconstruct the input. As diffusion model is trained on ID images, it is expected for the model to reconstruct it well in comparison to OOD images. The paper notes that simply this score is not enough for good OOD detection performance. Hence, the paper then measures how strongly the network’s intermediate representation reacts to slight noise nudge in the input. For ID samples, the network is expected to react less, and vice versa for OOD samples.

Critical strengths

The paper provides the insight the mere reconstruction cannot reliably be used for OOD detection. And it proposes to measure the intermediate representation's reaction to the perturbation in input space. And combine this scoring along with the reconstruction error to provide the single OOD detection score. The hyperparameters are determined by ID data only.

It provides a novel scoring method that outperforms the compared baselines. The performance gap with the best baseline is non-trivial. The paper complements the empirical performance with theoretical analyses too.

Critical weaknesses

For ImageNet near OOD, the authors are suggested to use SSB-Hard and NINCO. And for far OOD, please use: iNaturalist, Textures, OpenImage-O. Please see OpenOOD v1.5.

The results on FPR@95 metric is not reported. And, half of the benchmarks seems saturated with AUROC scores being near 1.0?

The OOD tuning dataset is CelebA when it is also used as test OOD set. The authors are suggested to not use the same dataset for tuning which is used for testing.

The paper doesn't report the results on transformers based architectures.

Does this method apply to flow matching as well?

Questions for authors

Is the image preprocessing pipeline identical for both ID and OOD images?

6 comments
AuthorsOpenReview ↗

Response to Reviewer DNMx (1/3)

We thank the reviewer for the positive assessment and the precise suggestions. We have uploaded a revised version of the paper, with all changes highlighted in blue, and address each comment point by point below.

W1. For ImageNet near OOD, the authors are suggested to use SSB-Hard and NINCO. And for far OOD, please use: iNaturalist, Textures, OpenImage-O. Please see OpenOOD v1.5.

Prompted by your comment, we extended the ImageNet-64 benchmark (App. F.3) with the suggested datasets.

MethodSVHNTexturesSSB-hardNINCOiNaturalistOpenImage-OAvg
DDPM-OOD0.2080.5700.5220.5510.8430.5950.548
LMD0.9010.4860.4800.4800.2160.3710.489
GEPC0.9650.7690.6240.6500.7480.7590.753
DiffPath0.7530.6990.6240.7190.8650.7420.734
DiffPathV20.5470.8600.7450.7970.8390.7500.756
EigenScore0.7550.7090.5410.4880.6020.5750.612
DUO (ours)0.9520.8130.7680.7600.6180.7240.773

DUO attains the highest average AUROC over all six OOD sets.

AuthorsOpenReview ↗

Response to Reviewer DNMx (2/3)

W2. The results on FPR@95 metric is not reported. And, half of the benchmarks seems saturated with AUROC scores being near 1.0?

Prompted by your comment, we now report FPR95, together with AUPR-In/Out and detection error, for the main (now in App. F.1), near-OOD and high-resolution benchmarks. The main benchmark with AUROC metric is the standard in prior diffusion-based OOD detection [1–4]; our evaluation goes beyond it in both benchmarks and metrics.

FPR95 ↓

MethodCelebA→C10CelebA→C100CelebA→SVHNC10→C100C10→SVHNC10→CelebAC100→C10C100→CelebAC100→SVHNSVHN→C10SVHN→C100SVHN→CelebAAvg
SCOPED0.3280.3640.5160.9420.9960.6980.9520.7740.9900.7040.6740.2620.683
GEPC0.1980.2640.1600.9020.8960.3440.9820.9960.9500.4400.3900.4900.584
DiffPath0.0000.0000.1180.8700.7300.1080.9740.2200.8340.2980.2200.9560.444
EigenScore0.1720.2420.4480.5620.8180.6940.9120.9760.8860.0460.1000.0260.490
DUO (ours)0.0020.0060.0000.4260.0620.6240.8840.9640.5900.0040.0160.0000.298

AUPR-In ↑

MethodCelebA→C10CelebA→C100CelebA→SVHNC10→C100C10→SVHNC10→CelebAC100→C10C100→CelebAC100→SVHNSVHN→C10SVHN→C100SVHN→CelebAAvg
SCOPED0.8830.8560.8800.5330.6910.8740.4680.8440.6450.7670.7380.9240.759
GEPC0.9370.9270.9370.5670.7520.9400.4480.6720.7200.8340.8410.8510.786
DiffPath1.0001.0000.9570.5450.7690.9820.4750.9750.7390.9360.9480.5270.821
EigenScore0.9590.9300.8920.8840.8430.8980.6810.5150.7030.9930.9760.9940.856
DUO (ours)1.0000.9991.0000.8470.9870.8980.6640.6190.8360.9960.9950.9980.903

AUPR-Out ↑

MethodCelebA→C10CelebA→C100CelebA→SVHNC10→C100C10→SVHNC10→CelebAC100→C10C100→CelebAC100→SVHNSVHN→C10SVHN→C100SVHN→CelebAAvg
SCOPED0.8860.8790.8490.5360.5570.8210.4690.7800.5050.7970.7890.9390.734
GEPC0.9440.9290.9490.5840.6890.9310.4450.4850.6340.8840.8950.8710.770
DiffPath1.0001.0000.9770.5850.7330.9230.4620.9010.6680.9250.9480.4860.801
EigenScore0.9640.9490.8850.8610.7400.8170.6150.4470.6400.9920.9830.9910.824
DUO (ours)1.0000.9991.0000.8910.9860.8410.6350.5070.8380.9930.9920.9930.890

Detection error ↓

MethodCelebA→C10CelebA→C100CelebA→SVHNC10→C100C10→SVHNC10→CelebAC100→C10C100→CelebAC100→SVHNSVHN→C10SVHN→C100SVHN→CelebAAvg
SCOPED0.1660.1810.1960.4390.3050.2130.4970.2210.3600.2500.2810.1460.271
GEPC0.1090.1370.1030.4290.2970.1330.5000.3620.3010.2080.2060.2110.250
DiffPath0.0010.0010.0820.4300.2990.0470.4960.0600.3210.1390.1120.4690.205
EigenScore0.0950.1270.1980.2000.2510.1970.3790.4690.3780.0410.0670.0310.203
DUO (ours)0.0030.0070.0050.1840.0530.2010.3740.4210.2350.0180.0290.0080.128

High Resolution:

FPR95 ↓

MethodAFHQMicroscopyAvg
GEPC0.4540.4140.434
DUO (ours)0.0000.0000.000

AUPR-In ↑

MethodAFHQMicroscopyAvg
GEPC0.7580.8000.779
DUO (ours)0.9991.0000.999

AUPR-Out ↑

MethodAFHQMicroscopyAvg
GEPC0.8510.8670.859
DUO (ours)0.9991.0000.999

Detection error ↓

MethodAFHQMicroscopyAvg
GEPC0.2210.2180.220
DUO (ours)0.0130.0080.011
AuthorsOpenReview ↗

Response to Reviewer DNMx (3/3)

Near-OOD:

FPR95 ↓

MethodC10 → C100C10 → TINC100 → C10C100 → TINAvg
EigenScore0.5620.4740.9120.7500.674
DUO (ours)0.4260.2680.8840.7140.573

AUPR-In ↑

MethodC10 → C100C10 → TINC100 → C10C100 → TINAvg
EigenScore0.8840.8980.6810.7540.804
DUO (ours)0.8470.9560.6640.8810.837

AUPR-Out ↑

MethodC10 → C100C10 → TINC100 → C10C100 → TINAvg
EigenScore0.8610.8970.6150.7440.779
DUO (ours)0.8910.9450.6350.8260.824

Detection error ↓

MethodC10 → C100C10 → TINC100 → C10C100 → TINAvg
EigenScore0.2000.1700.3790.3100.265
DUO (ours)0.1840.1150.3740.2100.221

Note that DUO also has the best average AUPR-In, AUPR-Out and detection error on the main benchmark, and it outperforms the strongest baseline in every metric on near-OOD and high resolution.

W3. The OOD tuning dataset is CelebA when it is also used as test OOD set. The authors are suggested to not use the same dataset for tuning which is used for testing.

We clarify that DUO uses no OOD data for any tuning or selection; every choice is made from ID calibration data only. Moreover, its advantage also does not depend on these choices: with random calibration subsets as small as 25 images, or with the tap fixed and γ = 0.5, DUO still outperforms EigenScore on average (see our responses to Reviewer 7oZn's Design comment, Q5 and Q7). App. F.7 reports only a diagnostic variant when a limited OOD validation set is available; this variant is excluded from the other results.

W4. The paper doesn't report the results on transformers based architectures.

Prompted by your comment, we ran DUO on a pixel-space diffusion transformer (U-ViT-S/2 [5], CIFAR-10 as ID, now in App. F.6). DUO needs only a denoiser and an intermediate feature map, so no part of the readout depends on a U-Net. To the best of our knowledge, none of [1–4] reports transformer results. While we report GEPC with U-ViT-S/2, GEPC [3] assumes that convolutional backbones trained with augmentations yield approximately equivariant scores; on the U-ViT transformer, this premise need not hold, and GEPC drops from 0.756 to 0.581.

MethodC10 → C100C10 → SVHNC10 → CelebAAvg
U-ViT-S/2 (pixel-space diffusion transformer)
SCOPED0.482 / 0.9140.794 / 0.3580.952 / 0.1020.743 / 0.458
GEPC0.598 / 0.8600.700 / 0.9580.445 / 0.9960.581 / 0.938
DiffPath0.646 / 0.8440.730 / 0.8940.919 / 0.2760.765 / 0.671
EigenScore0.846 / 0.7040.767 / 0.9020.833 / 0.8080.815 / 0.805
DUO0.748 / 0.7500.841 / 0.5600.872 / 0.7480.820 / 0.686

W5. Does this method apply to flow matching as well?

Prompted by your comment, we applied DUO unchanged to an OT-CFM flow-matching model [6] (CIFAR-10 as ID, now in App. F.6), using the denoiser given by its velocity field, x̂₁ = x_t + (1 − t)·v(x_t, t):

MethodC10 → C100C10 → SVHNC10 → CelebAAvg
OT-CFM (flow matching)
SCOPED0.552 / 0.8940.898 / 0.3300.968 / 0.1020.806 / 0.442
GEPC0.545 / 0.9040.661 / 0.9720.362 / 0.9920.522 / 0.956
DiffPath0.598 / 0.8640.955 / 0.2160.999 / 0.0000.851 / 0.360
EigenScore0.822 / 0.6480.621 / 0.9600.816 / 0.8260.753 / 0.811
DUO0.813 / 0.5640.902 / 0.3900.928 / 0.3120.881 / 0.422

DUO achieves the highest average AUROC.

Q1. Is the image preprocessing pipeline identical for both ID and OOD images?

We clarify that ID and OOD images go through the identical preprocessing pipeline, following the evaluation protocol of EigenScore [1], detailed in App. D.1.

References

[1] EigenScore: OOD detection using posterior covariance in diffusion models. ICLR 2026.

[2] Out-of-distribution detection with a single unconditional diffusion model. NeurIPS 2024.

[3] GEPC: Group-equivariant posterior consistency for out-of-distribution detection in diffusion models. ICML 2026.

[4] SCOPED: Score–curvature out-of-distribution proximity evaluator for diffusion. ICLR 2026.

[5] All are Worth Words: A ViT Backbone for Diffusion Models. CVPR 2023.

[6] Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport. TMLR 2024

Reviewer_3BHUOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work provides a novel lens to the Out-Of-Distribution (OOD) detection task through the Diffusion Models (DMs). The key insight is that the intermediate features in the diffusion processes contain beneficial information that promotes the separation between In-Distribution (ID) data and OOD data. To be specific, the authors apply perturbations to the intermediate features, compute the covariance, and take the top eigenvalues as the sensitivity score. This sensitivity score is finally fused with the common reconstruction error as the final detection score.

Critical strengths
  1. It is encouraging to explore OOD detection with the help of the diffusion models.
  2. There are many theoretical results to support the methodology. But I do not check the mathematical details.
  3. Experiment results show the general advantageous OOD detection performance of the proposed method over other diffusion-based detection approaches.
Critical weaknesses
  1. The rationale behind DUO is not elaborated convincingly enough. This work does not sufficiently explain, either intuitively or through direct numerical evidence, why the intermediate-feature spectrum under perturbations should be a reliable ID-OOD separation signal. Note that Theorem 1 only provide a simplified theoretical result for a Bayes denoiser, which is insufficient to justify its application to trained diffusion models.

  2. The main experiments are largely restricted to 32×32 datasets. Although the paper includes FFHQ-256 (Appendix E.2) and ImageNet-64 (Appendix E.3), the standard and more challenging ImageNet-1k benchmark is missing, which limits the generality of the conclusions given ImageNet-1k’s much larger category diversity and resolution. In addition, the evaluation uses only 500 ID and 500 OOD images per pair (Appendix D.1) rather than the full test sets. This subsampling can substantially affect the reported AUROC, and the paper does not justify why full test sets were not used or report confidence intervals/variance for the comparison. The authors should clarify whether all baselines are evaluated on exactly the same 500/500 subsets, and provide full-set results or statistical uncertainty to support the claimed improvements.

  3. Regarding the reconstruction-based OOD detection topic, some related works leverage the classical PCA tool to explore the principal subspaces in features/gradients [a,b] to capture the ID-OOD separability, which differs from the encoder-decoder-based reconstruction. The authors are encouraged to include these works [a,b] into references for an enriched literature review. Including comparison results is further appreciated.

  4. I would like to raise a broader point for discussion. At present, popular highly capable OOD detectors are built entirely upon discriminative classifiers, delivering both competitive performance and low overhead—most notably those grounded in feature shaping. Against this backdrop, what added value remains in resorting to generative models for OOD detection? Developing and serving generative models is generally costlier and more demanding than working with discriminative ones. The authors should articulate where this work sits and what it uniquely offers.

[a] Kernel PCA for Out-of-Distribution Detection. NeurIPS 2024.
[b] GradPCA: Leveraging NTK Alignment for Reliable Out-of-Distribution Detection. ICLR 2026.

Questions for authors
  1. Could the authors further justify the detection rationale behind DUO, particularly on why perturbed feature sensitivity as an effective ID-OOD separation signal?
  2. Could the authors provide results on the ImageNet-1K dataset with full images in evaluation? As this method is training free, its computational cost is not very heavy.
  3. Could the authors further justify where this work sits and what it uniquely offers, particularly when compared with those classifier-based detection methods with both superior effectiveness and efficiency.
6 comments
AuthorsOpenReview ↗

Response to Reviewer 3BHU (1/4)

We thank the reviewer for the careful review and constructive suggestions. We have uploaded a revised version of the paper, with all changes highlighted in blue, and address each comment point by point below.

W1. The rationale behind DUO is not elaborated convincingly enough. This work does not sufficiently explain, either intuitively or through direct numerical evidence, why the intermediate-feature spectrum under perturbations should be a reliable ID-OOD separation signal. Note that Theorem 1 only provide a simplified theoretical result for a Bayes denoiser, which is insufficient to justify its application to trained diffusion models.

We clarify DUO’s rationale below and have revised Sec. 4 accordingly.

Intuition. When a noisy input could have come from several ID components, the denoiser predicts their average, so an input lying between components can be reconstructed well even though the model cannot tell which component generated it (Sec. 4.1, Fig. 3). OOD inputs that fall between ID components therefore evade reconstruction, while the posterior covariance still registers the ambiguity (Thm. 1). In these regions the model's belief about the component changes quickly with the input, so a small perturbation moves intermediate features strongly and along a few directions. Prop. 4 proves this in the model, and Fig. 7 confirms it in the trained U-Nets. DUO's score measures this response through the largest eigenvalues of the feature-response covariance over random probes (Eq. 8).

Evidence. In trained U-Nets, reconstruction fails where OOD data are easier to denoise (App. B.2), and the intermediate tap keeps the signal that the output loses (sensitivity channel, AUROC):

Pair (ID → OOD)Selected tapDenoiser output
CIFAR-10 → CIFAR-1000.8610.607
CIFAR-10 → SVHN0.9880.141
CIFAR-100 → CIFAR-100.6330.547
CIFAR-100 → SVHN0.8990.147

Please see further numerical evidence on trained U-Nets in App. B.3.

We clarify that Theorem 1 is not meant to justify DUO on trained denoisers by itself; the evidence above does that empirically. Theorem 1 shows, in a tractable model, that the posterior covariance stays informative where reconstruction inverts, motivating sensitivity beyond reconstruction (Sec. 4.1). App. C.2 extends the result beyond two sheets to mixtures of low-rank Gaussians, a tractable model that captures the low-dimensional structure of real image data [12], and shows that spectral scoring stays informative where reconstruction inverts (Fig. 10). App. C.4 then shows that a feature can retain this ambiguity while the output loses it, even for a near-optimal denoiser (Prop. 4), which is why DUO reads an intermediate feature.

W2. The main experiments are largely restricted to 32×32 datasets. Although the paper includes FFHQ-256 (Appendix E.2) and ImageNet-64 (Appendix E.3), the standard and more challenging ImageNet-1k benchmark is missing, which limits the generality of the conclusions given ImageNet-1k’s much larger category diversity and resolution.

Prompted by your comment, we evaluated DUO on the ImageNet-1K benchmark following the protocol of [b] (iNaturalist, SUN, Places and Textures as OOD), using a 256×256 unconditional diffusion backbone shared by all methods (now in App. F.5). To our knowledge, none of [1–4] reports this benchmark.

Method (AUROC / FPR95)iNaturalistSUNPlacesTexturesAvg
DiffPath0.573 / 0.9440.598 / 0.9360.702 / 0.9060.693 / 0.8520.641 / 0.909
GEPC0.404 / 0.9800.489 / 0.9040.318 / 0.9780.423 / 0.9560.408 / 0.955
SCOPED0.719 / 0.7840.695 / 0.8180.800 / 0.8300.659 / 0.8400.718 / 0.818
EigenScore0.428 / 0.9440.651 / 0.8520.559 / 0.9010.630 / 0.8580.567 / 0.889
DUO0.756 / 0.8000.781 / 0.7320.882 / 0.6200.880 / 0.4240.825 / 0.644

DUO achieves the best AUROC on every OOD set and the lowest average FPR95, so its gains extend to ImageNet-1K’s category diversity and resolution. Together with FFHQ-256, ImageNet-64 (now extended with the OpenOOD v1.5 sets) and ImageNet-1K, the evaluation covers resolutions from 32×32 to 256×256.

AuthorsOpenReview ↗

Response to Reviewer 3BHU (2/4)

The authors should clarify whether all baselines are evaluated on exactly the same 500/500 subsets, and provide full-set results or statistical uncertainty to support the claimed improvements.

All baselines are evaluated on exactly the same 500/500 images, using the released split of [1].

Moreover, Table 21 (App. F.12) reports the standard deviation over 5 seeds on the main benchmark; the standard deviation of the average AUROC is 0.005.

MetricCelebACIFAR-10CIFAR-100SVHNAvg
AUROC0.9996 ± 0.00040.9113 ± 0.00440.6804 ± 0.01720.9946 ± 0.00490.8965 ± 0.0051

Prompted by your comment, we also report the seed sensitivity of the other metrics over 5 seeds (now in App. F.12):

MetricCelebACIFAR-10CIFAR-100SVHNAvg
FPR95 ↓0.0027 ± 0.00160.3707 ± 0.01820.8127 ± 0.03750.0067 ± 0.01070.2982 ± 0.0108
AUPR-In ↑0.9997 ± 0.00050.9106 ± 0.00870.7065 ± 0.01540.9965 ± 0.00310.9033 ± 0.0048
AUPR-Out ↑0.9997 ± 0.00030.9061 ± 0.00220.6598 ± 0.01830.9928 ± 0.01310.8896 ± 0.0067
Det. err ↓0.0050 ± 0.00260.1460 ± 0.00610.3433 ± 0.01230.0183 ± 0.00570.1282 ± 0.0041

Prompted by your comment, we re-evaluated DUO and the closest diffusion-based baselines on the full test sets (now in App. G). Note that DUO achieves the best performance.

AUROC ↑

MethodCelebA→C10CelebA→C100CelebA→SVHNC10→C100C10→SVHNC10→CelebAC100→C10C100→CelebAC100→SVHNSVHN→C10SVHN→C100SVHN→CelebAAvg
SCOPED0.8780.8880.9090.5600.6320.8140.4460.7710.5340.8020.7860.8620.740
GEPC0.9440.9260.9630.5720.8350.9830.4340.6010.7920.8610.8760.9180.809
DiffPath1.0001.0000.9390.5670.8560.9140.4380.9170.8350.9390.9510.4360.816
EigenScore0.9560.9360.8490.8660.7940.8320.6380.4000.6430.9890.9770.9870.822
DUO1.0001.0001.0000.8710.9850.8930.6660.4670.8380.9940.9930.9940.892

FPR95 ↓

MethodCelebA→C10CelebA→C100CelebA→SVHNC10→C100C10→SVHNC10→CelebAC100→C10C100→CelebAC100→SVHNSVHN→C10SVHN→C100SVHN→CelebAAvg
SCOPED0.4010.3900.4040.9330.9950.9340.9580.9240.9960.7010.6820.4110.727
GEPC0.2200.2640.1290.9000.8610.0940.9630.9820.9340.4390.3840.3440.543
DiffPath0.0000.0000.2180.8760.5730.0770.9730.0830.6470.3420.2680.9540.418
EigenScore0.2060.2720.5640.5890.8460.7870.9060.9820.8930.0520.1140.0610.523
DUO0.0020.0030.0040.4410.0670.5640.8860.9630.6050.0070.0150.0000.296
AuthorsOpenReview ↗

Response to Reviewer 3BHU (3/4)

W3. Regarding the reconstruction-based OOD detection topic, some related works leverage the classical PCA tool to explore the principal subspaces in features/gradients [a,b] to capture the ID-OOD separability, which differs from the encoder-decoder-based reconstruction. The authors are encouraged to include these works [a,b] into references for an enriched literature review. Including comparison results is further appreciated.

Prompted by your comment, we now discuss [a] and [b] and provide a comparison with both (now in App. E).

ID → OODCoP [a]CoRP [a]GradPCA [b]EigenScoreDUO
CelebA → C100.804 / 0.9320.837 / 0.9200.816 / 0.9420.965 / 0.1721.000 / 0.002
CelebA → C1000.802 / 0.9280.841 / 0.8860.808 / 0.9240.944 / 0.2421.000 / 0.006
CelebA → SVHN0.940 / 0.3800.941 / 0.5000.865 / 0.9740.888 / 0.4481.000 / 0.000
C10 → C1000.749 / 0.9680.938 / 0.2480.904 / 0.3440.880 / 0.5620.876 / 0.426
C10 → SVHN0.769 / 0.9940.958 / 0.2100.914 / 0.3580.810 / 0.8180.986 / 0.062
C10 → CelebA0.767 / 0.9840.943 / 0.2220.892 / 0.3840.873 / 0.6940.872 / 0.624
C100 → C100.553 / 0.9520.576 / 0.9800.587 / 0.9440.642 / 0.9120.657 / 0.884
C100 → CelebA0.769 / 0.8420.684 / 0.9060.804 / 0.9000.427 / 0.9760.544 / 0.964
C100 → SVHN0.778 / 0.7680.834 / 0.6900.659 / 0.9480.661 / 0.8860.839 / 0.590
SVHN → C100.777 / 0.9980.975 / 0.1180.956 / 0.2260.992 / 0.0460.995 / 0.004
SVHN → C1000.784 / 0.9940.975 / 0.1000.954 / 0.2340.982 / 0.1000.994 / 0.016
SVHN → CelebA0.785 / 1.0000.985 / 0.0700.944 / 0.3420.994 / 0.0260.996 / 0.000
Average0.773 / 0.8950.874 / 0.4870.842 / 0.6270.838 / 0.4900.897 / 0.298

W4. I would like to raise a broader point for discussion. At present, popular highly capable OOD detectors are built entirely upon discriminative classifiers, delivering both competitive performance and low overhead—most notably those grounded in feature shaping. Against this backdrop, what added value remains in resorting to generative models for OOD detection? Developing and serving generative models is generally costlier and more demanding than working with discriminative ones. The authors should articulate where this work sits and what it uniquely offers.

Classifier-based detectors judge inputs through a classifier trained on labeled ID data, so their notion of OOD is tied to the classes the classifier learned. A growing class of systems requires a different test: whether an input lies within the training distribution of a generative model itself. Diffusion models are now used as priors for inverse problems in medical and scientific imaging, as planners in reinforcement learning, and as robot policies [7, 8, 9]. There, an input outside the model’s training distribution can produce hallucinated reconstructions [7] or unreliable plans [17], so it must be detected with respect to that model’s own distribution. The same need arises whenever labels are unavailable, as in medical anomaly detection [10] and zero-shot anomaly detection [11]. Thus, unsupervised OOD detection with generative models is an established problem with its own benchmark. It started with likelihood-based detectors and their failure modes [13, 14], and diffusion-based detectors have followed since 2023 [15, 16, 2, 11], including EigenScore [1], SCOPED [4] and GEPC [3]. DUO follows this line of work (Table 1) and improves on it within the same setting, where neither labels nor OOD data are available and only a model of the ID data is at hand. Classifier-based detectors need labeled ID data, so they do not apply in this setting. We address the cost in our response to Q3. We have added this positioning to App. E.

[a] Kernel PCA for Out-of-Distribution Detection. NeurIPS 2024.

[b] GradPCA: Leveraging NTK Alignment for Reliable Out-of-Distribution Detection. ICLR 2026.

Q1. Could the authors further justify the detection rationale behind DUO, particularly on why perturbed feature sensitivity as an effective ID-OOD separation signal?

Please see our response to W1.

Q2. Could the authors provide results on the ImageNet-1K dataset with full images in evaluation? As this method is training free, its computational cost is not very heavy.

Prompted by your comment, we report DUO and the diffusion-based baselines on the ImageNet-1K benchmark in our response to W2.

Reviewer_7oZnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the use of a diffusion model trained on in-distribution (ID) data for out of distribution detection (OOD). The paper introduces a sensitivity spectrum, based on Intermediate features, which is shown to be complementary to the reconstruction error. The core idea is to perturb the input and check how sensitive an intermediate feature is, by comparing the feature responses under small perturbations along 32 random directions (finite difference). The reconstruction and sensitivity scores are standardized using a held-out ID calibration set, aggregated over selected noise levels, and combined through a weighted linear score (the sensitivity score is combined with the reconstruction error using a z-score-style, ID-calibrated fusion).

The experiments validate the approach with improved AUROC over the strongest baseline presented.

Critical strengths
  • Easy to understand, clear pipeline. The method reads out two channels from a frozen diffusion model (sensitivity and re-construction error) where the sensitivity signal is constructed by perturbing the input and measuring the response of an intermediate feature.
  • Intuitive description of the method
  • Implementation with significantly lower inference time computational cost compared to others using diffusion models for the task.
Critical weaknesses
  • Novelty claim. Compared to prior work, DUO seems to move the spectral readout from the output to an intermediate layer and add reconstruction to improve the overall score. Then EigenScore’s output Jacobian has a direct posterior-covariance interpretation but DUO’s intermediate-feature Jacobian generally does not. Fundamentally not a new signal given EigenScore method. Then, fusion amounts to a weighted sum. The two scores are fused by a z-score standardization followed by a weighted sum. No other fusion method is compared, so it is unclear whether the gain could be larger with a better fusion approach and which scores are helpful and under what context or for what samples or insights into the observed results.

  • Reporting. AUROC is useful for ranking, but it can conceal poor performance at operational thresholds. For OOD detection, metrics such as FPR95, AUPR-OOD, AUPR-ID, and detection error are very informative. The lack of greater emphasis on threshold-sensitive measures makes it difficult to assess whether the method's AUROC gains translate into operationally useful reductions in false alarms. Then, the principal table emphasizes point estimates, while the method will be impacted by randomness, just requires multiple evals and standard deviation. Further, several main benchmark pairs, such as natural-object datasets versus SVHN or CelebA, correspond to visually large distribution shifts. AUROC values near 1.0 for multiple methods indicate that some of these tasks are close to saturated. The near-OOD evaluation is therefore especially important, but much of it is moved to the appendix.

  • Design. It is not clear how the held-out calibration per dataset of ID-OOD eval is determined and the sensitivity of the method to this set as many parameters are selected based on this set to achieve the reported performance. Fusion is a weighted sum, based on the calibration, and this is a weakness as the calibration sensitivity is not known.

  • (Minor but important) Abstract suggest the paper is about ODD detection for Diffusion models as opposed to the use of Generative Models for OOD detection. Overall, the motivation for the method is poor, where claims of training free etc. ignore the cost of building an ID diffusion models.

Questions for authors
  1. What is the criteria for selecting gamma?
  2. On which ID-OOD pairs does sensitivity materially improve over reconstruction, and on which pairs does it hurt? Aggregate AUROC does not fully expose this behaviour.
  3. How do latency, memory, and accuracy scale with image resolution, feature-map dimensionality, probe count, and batch size?
  4. Are the strong results dependent on selecting transformations that accidentally resemble the evaluated OOD shifts?
  5. Could the authors report a proxy free baseline?
  6. Have you tried other ways of fusing the two scores, such as a small MLP on (z_R, z_J) trained to separate ID samples from their proxy-shifted (rotated/flipped) versions, which serve as OOD?
  7. How sensitive is DUO to ID calibration images used (e.g. 50 or 100)?
  8. Can the authors clearly articulate a novelty position in terms of existing knowledge in the field for OOD detection, other than estimation error alone can fail?
  9. It is not fully clear whether all baselines use, was the entire test sets of the datasets used for OOD or what was selected and how, the same diffusion architecture and training data, the same number of noise samples where applicable, equally careful hyperparameter selection, the same ID-only validation restrictions etc. Can such details be explicitly clarified?
  10. It would be useful to know whether the improvement over EigenScore holds consistently across seeds and across individual ID-OOD pairs.
  11. Were the inference times method fairly, i.e removing cold starts, data loads, a quiet machine etc. being considered?
7 comments
AuthorsOpenReview ↗

Response to Reviewer 7oZn (1/6)

We thank the reviewer for the careful review and constructive suggestions. We have uploaded a revised version of the paper, with all changes highlighted in blue, and address each comment point by point below.

Novelty claim. Compared to prior work, DUO seems to move the spectral readout from the output to an intermediate layer and add reconstruction to improve the overall score. Then EigenScore’s output Jacobian has a direct posterior-covariance interpretation but DUO’s intermediate-feature Jacobian generally does not. Fundamentally not a new signal given EigenScore method. Then, fusion amounts to a weighted sum.

DUO’s main novelty is reading the denoiser’s intermediate response for OOD detection, backed by an analysis of why output-level readouts fail. We show when output-level detectors fail, from reconstruction error to the output covariance of EigenScore [1], explain why, and prescribe where to read instead and how to use it. Moving the readout is not cosmetic; it turns an inverted detector into a reliable one. The output covariance fails when the readout discards uncertainty that intermediate features keep (Prop. 4). The same estimator, read at both locations, shows this (sensitivity channel, AUROC; details in App. B.3):

Pair (ID → OOD)Selected tapDenoiser output
CIFAR-10 → CIFAR-1000.8610.607
CIFAR-10 → SVHN0.9880.141
CIFAR-100 → CIFAR-100.6330.547
CIFAR-100 → SVHN0.8990.147

Reconstruction is not added to raise the score. Sensitivity alone already outperforms EigenScore (Table 19). Reconstruction is added because the two channels fail in different places (Prop. 1, Q2) and are negatively correlated on ID data (Table 5). Please see further numerical support in App. B.4.

DUO’s feature sensitivity has an interpretation. It measures the same branch uncertainty as the posterior covariance (Prop. 4). EigenScore’s insight that posterior uncertainty signals OOD is sound, and our Theorem 1 confirms it; DUO shows that uncertainty is best read at intermediate features.

We have reordered the contributions in Sec. 1 to lead with this diagnosis.

The two scores are fused by a z-score standardization followed by a weighted sum. No other fusion method is compared, so it is unclear whether the gain could be larger with a better fusion approach and which scores are helpful and under what context or for what samples or insights into the observed results.

Prompted by your comment, we compared alternative fusions, including the proposed MLP (Q6): DUO beats all of them. DUO’s fusion is principled, not ad hoc: Thm. 2 proves why it works and when, and App. B.4 shows that this model tracks the measured AUROC across fusion weights. Q2 shows which channel carries each pair.

Reporting. AUROC is useful for ranking, but it can conceal poor performance at operational thresholds. For OOD detection, metrics such as FPR95, AUPR-OOD, AUPR-ID, and detection error are very informative. The lack of greater emphasis on threshold-sensitive measures makes it difficult to assess whether the method's AUROC gains translate into operationally useful reductions in false alarms.

Prompted by your comment, we now report FPR95, AUPR-In, AUPR-Out and detection error across our experiments (now included in App. F.1): the main 12-pair benchmark, near-OOD and high resolution. AUROC on this benchmark is the standard metric in prior diffusion-based OOD detection [1–4]; our evaluation now goes beyond it in both benchmarks and metrics.

FPR95 ↓

MethodCelebA→C10CelebA→C100CelebA→SVHNC10→C100C10→SVHNC10→CelebAC100→C10C100→CelebAC100→SVHNSVHN→C10SVHN→C100SVHN→CelebAAvg
SCOPED0.3280.3640.5160.9420.9960.6980.9520.7740.9900.7040.6740.2620.683
GEPC0.1980.2640.1600.9020.8960.3440.9820.9960.9500.4400.3900.4900.584
DiffPath0.0000.0000.1180.8700.7300.1080.9740.2200.8340.2980.2200.9560.444
EigenScore0.1720.2420.4480.5620.8180.6940.9120.9760.8860.0460.1000.0260.490
DUO (ours)0.0020.0060.0000.4260.0620.6240.8840.9640.5900.0040.0160.0000.298
AuthorsOpenReview ↗

Response to Reviewer 7oZn (2/6)

AUPR-In ↑

MethodCelebA→C10CelebA→C100CelebA→SVHNC10→C100C10→SVHNC10→CelebAC100→C10C100→CelebAC100→SVHNSVHN→C10SVHN→C100SVHN→CelebAAvg
SCOPED0.8830.8560.8800.5330.6910.8740.4680.8440.6450.7670.7380.9240.759
GEPC0.9370.9270.9370.5670.7520.9400.4480.6720.7200.8340.8410.8510.786
DiffPath1.0001.0000.9570.5450.7690.9820.4750.9750.7390.9360.9480.5270.821
EigenScore0.9590.9300.8920.8840.8430.8980.6810.5150.7030.9930.9760.9940.856
DUO (ours)1.0000.9991.0000.8470.9870.8980.6640.6190.8360.9960.9950.9980.903

AUPR-Out ↑

MethodCelebA→C10CelebA→C100CelebA→SVHNC10→C100C10→SVHNC10→CelebAC100→C10C100→CelebAC100→SVHNSVHN→C10SVHN→C100SVHN→CelebAAvg
SCOPED0.8860.8790.8490.5360.5570.8210.4690.7800.5050.7970.7890.9390.734
GEPC0.9440.9290.9490.5840.6890.9310.4450.4850.6340.8840.8950.8710.770
DiffPath1.0001.0000.9770.5850.7330.9230.4620.9010.6680.9250.9480.4860.801
EigenScore0.9640.9490.8850.8610.7400.8170.6150.4470.6400.9920.9830.9910.824
DUO (ours)1.0000.9991.0000.8910.9860.8410.6350.5070.8380.9930.9920.9930.890

Detection error ↓

MethodCelebA→C10CelebA→C100CelebA→SVHNC10→C100C10→SVHNC10→CelebAC100→C10C100→CelebAC100→SVHNSVHN→C10SVHN→C100SVHN→CelebAAvg
SCOPED0.1660.1810.1960.4390.3050.2130.4970.2210.3600.2500.2810.1460.271
GEPC0.1090.1370.1030.4290.2970.1330.5000.3620.3010.2080.2060.2110.250
DiffPath0.0010.0010.0820.4300.2990.0470.4960.0600.3210.1390.1120.4690.205
EigenScore0.0950.1270.1980.2000.2510.1970.3790.4690.3780.0410.0670.0310.203
DUO (ours)0.0030.0070.0050.1840.0530.2010.3740.4210.2350.0180.0290.0080.128

High Resolution:

FPR95 ↓

MethodAFHQMicroscopyAvg
GEPC0.4540.4140.434
DUO (ours)0.0000.0000.000

AUPR-In ↑

MethodAFHQMicroscopyAvg
GEPC0.7580.8000.779
DUO (ours)0.9991.0000.999

AUPR-Out ↑

MethodAFHQMicroscopyAvg
GEPC0.8510.8670.859
DUO (ours)0.9991.0000.999

Detection error ↓

MethodAFHQMicroscopyAvg
GEPC0.2210.2180.220
DUO (ours)0.0130.0080.011

Near-OOD:

FPR95 ↓

MethodC10 → C100C10 → TINC100 → C10C100 → TINAvg
EigenScore0.5620.4740.9120.7500.674
DUO (ours)0.4260.2680.8840.7140.573

AUPR-In ↑

MethodC10 → C100C10 → TINC100 → C10C100 → TINAvg
EigenScore0.8840.8980.6810.7540.804
DUO (ours)0.8470.9560.6640.8810.837

AUPR-Out ↑

MethodC10 → C100C10 → TINC100 → C10C100 → TINAvg
EigenScore0.8610.8970.6150.7440.779
DUO (ours)0.8910.9450.6350.8260.824

Detection error ↓

MethodC10 → C100C10 → TINC100 → C10C100 → TINAvg
EigenScore0.2000.1700.3790.3100.265
DUO (ours)0.1840.1150.3740.2100.221

Note that DUO also has the best average AUPR-In, AUPR-Out and detection error on the main benchmark, and it outperforms the strongest baseline in every metric on near-OOD and high resolution.

AuthorsOpenReview ↗

Response to Reviewer 7oZn (3/6)

Then, the principal table emphasizes point estimates, while the method will be impacted by randomness, just requires multiple evals and standard deviation.

Table 21 (App. F.12) reports the standard deviation over 5 seeds on the main benchmark; the standard deviation of the average AUROC is 0.005.

MetricCelebACIFAR-10CIFAR-100SVHNAvg
AUROC0.9996 ± 0.00040.9113 ± 0.00440.6804 ± 0.01720.9946 ± 0.00490.8965 ± 0.0051

Prompted by your comment, we provided the results on the full test set (now in App. G).

AUROC ↑

MethodCelebA→C10CelebA→C100CelebA→SVHNC10→C100C10→SVHNC10→CelebAC100→C10C100→CelebAC100→SVHNSVHN→C10SVHN→C100SVHN→CelebAAvg
SCOPED0.8780.8880.9090.5600.6320.8140.4460.7710.5340.8020.7860.8620.740
GEPC0.9440.9260.9630.5720.8350.9830.4340.6010.7920.8610.8760.9180.809
DiffPath1.0001.0000.9390.5670.8560.9140.4380.9170.8350.9390.9510.4360.816
EigenScore0.9560.9360.8490.8660.7940.8320.6380.4000.6430.9890.9770.9870.822
DUO1.0001.0001.0000.8710.9850.8930.6660.4670.8380.9940.9930.9940.892

FPR95 ↓

MethodCelebA→C10CelebA→C100CelebA→SVHNC10→C100C10→SVHNC10→CelebAC100→C10C100→CelebAC100→SVHNSVHN→C10SVHN→C100SVHN→CelebAAvg
SCOPED0.4010.3900.4040.9330.9950.9340.9580.9240.9960.7010.6820.4110.727
GEPC0.2200.2640.1290.9000.8610.0940.9630.9820.9340.4390.3840.3440.543
DiffPath0.0000.0000.2180.8760.5730.0770.9730.0830.6470.3420.2680.9540.418
EigenScore0.2060.2720.5640.5890.8460.7870.9060.9820.8930.0520.1140.0610.523
DUO0.0020.0030.0040.4410.0670.5640.8860.9630.6050.0070.0150.0000.296

Prompted by your comment, we also report the seed sensitivity of the other metrics over 5 seeds (now in App. F.12):

MetricCelebACIFAR-10CIFAR-100SVHNAvg
FPR95 ↓0.0027 ± 0.00160.3707 ± 0.01820.8127 ± 0.03750.0067 ± 0.01070.2982 ± 0.0108
AUPR-In ↑0.9997 ± 0.00050.9106 ± 0.00870.7065 ± 0.01540.9965 ± 0.00310.9033 ± 0.0048
AUPR-Out ↑0.9997 ± 0.00030.9061 ± 0.00220.6598 ± 0.01830.9928 ± 0.01310.8896 ± 0.0067
Det. err ↓0.0050 ± 0.00260.1460 ± 0.00610.3433 ± 0.01230.0183 ± 0.00570.1282 ± 0.0041

Further, several main benchmark pairs, such as natural-object datasets versus SVHN or CelebA, correspond to visually large distribution shifts. AUROC values near 1.0 for multiple methods indicate that some of these tasks are close to saturated. The near-OOD evaluation is therefore especially important, but much of it is moved to the appendix.

Prompted by your comment, we moved the near-OOD table to the main text. Note that DUO achieves the best average AUROC on near-OOD and in every threshold metric.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the rebuttal

We thank the AC and all reviewers for taking the time to review our paper and provide feedback. Here is the summary of our rebuttal (changes are highlighted in blue in the revised PDF).

  • Beyond the standard 12-pair benchmark [1–4], our initial submission already covered near-OOD, ImageNet-64, high-resolution and RL planning (D4RL) settings. In the revision, we extended the evaluation to ImageNet-1K (Reviewer 3BHU; App. F.5), the full test sets of all 12 pairs (Reviewers 3BHU and 7oZn; App. G), and the OpenOOD v1.5 sets with ImageNet-64 as ID (Reviewer DNMx; App. F.3).
  • We also added FPR95, AUPR-In, AUPR-Out and detection error metrics (Reviewers 7oZn and DNMx; App. F.1), where DUO has the best average on all four metrics.
  • Beyond the existing seed and sensitivity analyses (Reviewer 7oZn; App. F.9–F.12), we added the sensitivity to the calibration set and the proxy transformations (App. F.13).
  • Beyond our fusion weight, which is backed by the two-channel analysis (Sec. 4.3; Thm. 2, App. A.3), we compare with other fusion methods like an MLP (Reviewer 7oZn; App. F.11), showing that our principled fusion performs best.
  • We added a comparison with classifier-based OOD detectors in App. E (Reviewer 3BHU).
  • We added further scalability and cost analyses in App. H (Reviewer 7oZn).
  • We demonstrated the extension of DUO to diffusion transformers and flow matching (Reviewer DNMx; App. F.6).
  • We further clarified DUO's novelty and rationale (Reviewers 7oZn and 3BHU; Secs. 1 and 4.2, App. B.3). To the best of our knowledge, DUO is the first diffusion-based detector to measure how the denoiser's intermediate features respond to input perturbations.

We are glad that Reviewer 3BHU found all of their concerns addressed and raised their rating, and that Reviewer DNMx maintained their positive rating. Reviewer 7oZn has not responded to our rebuttal, but we believe our responses and the revised paper address each of their concerns.

References

[1] EigenScore: OOD detection using posterior covariance in diffusion models. ICLR 2026.

[2] Out-of-distribution detection with a single unconditional diffusion model. NeurIPS 2024.

[3] GEPC: Group-equivariant posterior consistency for out-of-distribution detection in diffusion models. ICML 2026.

[4] SCOPED: Score–curvature out-of-distribution proximity evaluator for diffusion. ICLR 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.