- 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
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 ↓
| Method | CelebA→C10 | CelebA→C100 | CelebA→SVHN | C10→C100 | C10→SVHN | C10→CelebA | C100→C10 | C100→CelebA | C100→SVHN | SVHN→C10 | SVHN→C100 | SVHN→CelebA | Avg |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SCOPED | 0.328 | 0.364 | 0.516 | 0.942 | 0.996 | 0.698 | 0.952 | 0.774 | 0.990 | 0.704 | 0.674 | 0.262 | 0.683 |
| GEPC | 0.198 | 0.264 | 0.160 | 0.902 | 0.896 | 0.344 | 0.982 | 0.996 | 0.950 | 0.440 | 0.390 | 0.490 | 0.584 |
| DiffPath | 0.000 | 0.000 | 0.118 | 0.870 | 0.730 | 0.108 | 0.974 | 0.220 | 0.834 | 0.298 | 0.220 | 0.956 | 0.444 |
| EigenScore | 0.172 | 0.242 | 0.448 | 0.562 | 0.818 | 0.694 | 0.912 | 0.976 | 0.886 | 0.046 | 0.100 | 0.026 | 0.490 |
| DUO (ours) | 0.002 | 0.006 | 0.000 | 0.426 | 0.062 | 0.624 | 0.884 | 0.964 | 0.590 | 0.004 | 0.016 | 0.000 | 0.298 |
AUPR-In ↑
| Method | CelebA→C10 | CelebA→C100 | CelebA→SVHN | C10→C100 | C10→SVHN | C10→CelebA | C100→C10 | C100→CelebA | C100→SVHN | SVHN→C10 | SVHN→C100 | SVHN→CelebA | Avg |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SCOPED | 0.883 | 0.856 | 0.880 | 0.533 | 0.691 | 0.874 | 0.468 | 0.844 | 0.645 | 0.767 | 0.738 | 0.924 | 0.759 |
| GEPC | 0.937 | 0.927 | 0.937 | 0.567 | 0.752 | 0.940 | 0.448 | 0.672 | 0.720 | 0.834 | 0.841 | 0.851 | 0.786 |
| DiffPath | 1.000 | 1.000 | 0.957 | 0.545 | 0.769 | 0.982 | 0.475 | 0.975 | 0.739 | 0.936 | 0.948 | 0.527 | 0.821 |
| EigenScore | 0.959 | 0.930 | 0.892 | 0.884 | 0.843 | 0.898 | 0.681 | 0.515 | 0.703 | 0.993 | 0.976 | 0.994 | 0.856 |
| DUO (ours) | 1.000 | 0.999 | 1.000 | 0.847 | 0.987 | 0.898 | 0.664 | 0.619 | 0.836 | 0.996 | 0.995 | 0.998 | 0.903 |
AUPR-Out ↑
| Method | CelebA→C10 | CelebA→C100 | CelebA→SVHN | C10→C100 | C10→SVHN | C10→CelebA | C100→C10 | C100→CelebA | C100→SVHN | SVHN→C10 | SVHN→C100 | SVHN→CelebA | Avg |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SCOPED | 0.886 | 0.879 | 0.849 | 0.536 | 0.557 | 0.821 | 0.469 | 0.780 | 0.505 | 0.797 | 0.789 | 0.939 | 0.734 |
| GEPC | 0.944 | 0.929 | 0.949 | 0.584 | 0.689 | 0.931 | 0.445 | 0.485 | 0.634 | 0.884 | 0.895 | 0.871 | 0.770 |
| DiffPath | 1.000 | 1.000 | 0.977 | 0.585 | 0.733 | 0.923 | 0.462 | 0.901 | 0.668 | 0.925 | 0.948 | 0.486 | 0.801 |
| EigenScore | 0.964 | 0.949 | 0.885 | 0.861 | 0.740 | 0.817 | 0.615 | 0.447 | 0.640 | 0.992 | 0.983 | 0.991 | 0.824 |
| DUO (ours) | 1.000 | 0.999 | 1.000 | 0.891 | 0.986 | 0.841 | 0.635 | 0.507 | 0.838 | 0.993 | 0.992 | 0.993 | 0.890 |
Detection error ↓
| Method | CelebA→C10 | CelebA→C100 | CelebA→SVHN | C10→C100 | C10→SVHN | C10→CelebA | C100→C10 | C100→CelebA | C100→SVHN | SVHN→C10 | SVHN→C100 | SVHN→CelebA | Avg |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SCOPED | 0.166 | 0.181 | 0.196 | 0.439 | 0.305 | 0.213 | 0.497 | 0.221 | 0.360 | 0.250 | 0.281 | 0.146 | 0.271 |
| GEPC | 0.109 | 0.137 | 0.103 | 0.429 | 0.297 | 0.133 | 0.500 | 0.362 | 0.301 | 0.208 | 0.206 | 0.211 | 0.250 |
| DiffPath | 0.001 | 0.001 | 0.082 | 0.430 | 0.299 | 0.047 | 0.496 | 0.060 | 0.321 | 0.139 | 0.112 | 0.469 | 0.205 |
| EigenScore | 0.095 | 0.127 | 0.198 | 0.200 | 0.251 | 0.197 | 0.379 | 0.469 | 0.378 | 0.041 | 0.067 | 0.031 | 0.203 |
| DUO (ours) | 0.003 | 0.007 | 0.005 | 0.184 | 0.053 | 0.201 | 0.374 | 0.421 | 0.235 | 0.018 | 0.029 | 0.008 | 0.128 |
High Resolution:
FPR95 ↓
| Method | AFHQ | Microscopy | Avg |
|---|---|---|---|
| GEPC | 0.454 | 0.414 | 0.434 |
| DUO (ours) | 0.000 | 0.000 | 0.000 |
AUPR-In ↑
| Method | AFHQ | Microscopy | Avg |
|---|---|---|---|
| GEPC | 0.758 | 0.800 | 0.779 |
| DUO (ours) | 0.999 | 1.000 | 0.999 |
AUPR-Out ↑
| Method | AFHQ | Microscopy | Avg |
|---|---|---|---|
| GEPC | 0.851 | 0.867 | 0.859 |
| DUO (ours) | 0.999 | 1.000 | 0.999 |
Detection error ↓
| Method | AFHQ | Microscopy | Avg |
|---|---|---|---|
| GEPC | 0.221 | 0.218 | 0.220 |
| DUO (ours) | 0.013 | 0.008 | 0.011 |
Response to Reviewer DNMx (3/3)
Near-OOD:
FPR95 ↓
| Method | C10 → C100 | C10 → TIN | C100 → C10 | C100 → TIN | Avg |
|---|---|---|---|---|---|
| EigenScore | 0.562 | 0.474 | 0.912 | 0.750 | 0.674 |
| DUO (ours) | 0.426 | 0.268 | 0.884 | 0.714 | 0.573 |
AUPR-In ↑
| Method | C10 → C100 | C10 → TIN | C100 → C10 | C100 → TIN | Avg |
|---|---|---|---|---|---|
| EigenScore | 0.884 | 0.898 | 0.681 | 0.754 | 0.804 |
| DUO (ours) | 0.847 | 0.956 | 0.664 | 0.881 | 0.837 |
AUPR-Out ↑
| Method | C10 → C100 | C10 → TIN | C100 → C10 | C100 → TIN | Avg |
|---|---|---|---|---|---|
| EigenScore | 0.861 | 0.897 | 0.615 | 0.744 | 0.779 |
| DUO (ours) | 0.891 | 0.945 | 0.635 | 0.826 | 0.824 |
Detection error ↓
| Method | C10 → C100 | C10 → TIN | C100 → C10 | C100 → TIN | Avg |
|---|---|---|---|---|---|
| EigenScore | 0.200 | 0.170 | 0.379 | 0.310 | 0.265 |
| DUO (ours) | 0.184 | 0.115 | 0.374 | 0.210 | 0.221 |
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.
| Method | C10 → C100 | C10 → SVHN | C10 → CelebA | Avg |
|---|---|---|---|---|
| U-ViT-S/2 (pixel-space diffusion transformer) | ||||
| SCOPED | 0.482 / 0.914 | 0.794 / 0.358 | 0.952 / 0.102 | 0.743 / 0.458 |
| GEPC | 0.598 / 0.860 | 0.700 / 0.958 | 0.445 / 0.996 | 0.581 / 0.938 |
| DiffPath | 0.646 / 0.844 | 0.730 / 0.894 | 0.919 / 0.276 | 0.765 / 0.671 |
| EigenScore | 0.846 / 0.704 | 0.767 / 0.902 | 0.833 / 0.808 | 0.815 / 0.805 |
| DUO | 0.748 / 0.750 | 0.841 / 0.560 | 0.872 / 0.748 | 0.820 / 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):
| Method | C10 → C100 | C10 → SVHN | C10 → CelebA | Avg |
|---|---|---|---|---|
| OT-CFM (flow matching) | ||||
| SCOPED | 0.552 / 0.894 | 0.898 / 0.330 | 0.968 / 0.102 | 0.806 / 0.442 |
| GEPC | 0.545 / 0.904 | 0.661 / 0.972 | 0.362 / 0.992 | 0.522 / 0.956 |
| DiffPath | 0.598 / 0.864 | 0.955 / 0.216 | 0.999 / 0.000 | 0.851 / 0.360 |
| EigenScore | 0.822 / 0.648 | 0.621 / 0.960 | 0.816 / 0.826 | 0.753 / 0.811 |
| DUO | 0.813 / 0.564 | 0.902 / 0.390 | 0.928 / 0.312 | 0.881 / 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
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.
Prompted by your comment, we extended the ImageNet-64 benchmark (App. F.3) with the suggested datasets.
DUO attains the highest average AUROC over all six OOD sets.