Diffusion OOD detection stops trusting reconstruction alone: DUO adds a feature-sensitivity readout and reaches 0.897 average AUROC on 12 pairs
Synopsis
Diffusion-model out-of-distribution detection often relies on reconstruction error, yet OOD inputs can be denoised just as well or better, so those scores overlap or even invert. DUO adds a second readout on a frozen diffusion model: the sensitivity spectrum of intermediate denoising features to small input perturbations, with the feature layer, noise levels and two-channel fusion calibrated on in-distribution data alone. Across 12 image pairs DUO reaches 0.897 average AUROC versus 0.838 for the strongest baseline, EigenScore, at 0.28 seconds per sample.
Interpretation
Out-of-distribution detection no longer asks only how accurately a diffusion model reconstructs an input; it reads reconstruction error together with the local sensitivity of intermediate denoising features to input perturbations, and fuses both into one one-sided score where larger values mean stronger OOD evidence. Earlier diffusion-based detectors mainly used denoising or reconstruction error and assumed in-distribution inputs are reconstructed more accurately; when OOD inputs share low-level statistics, texture or semantic structure with ID data, the two reconstruction scores overlap heavily and can even invert. On the 12 pairs formed by CIFAR-10, CIFAR-100, SVHN and CelebA against one another, DUO averages 0.897 AUROC against 0.838 for the strongest baseline EigenScore; on the FFHQ-256 high-resolution benchmark it averages 0.999 against 0.826 for GEPC.
The sensitivity channel never forms a full Jacobian; it probes the activation at one intermediate decoder layer of a noisy input with finite differences and takes the sum of the largest squared singular values of the response matrix as that layer and noise level's sensitivity statistic. The reconstruction channel measures the magnitude of the denoising residual, while the sensitivity channel measures how strongly intermediate features respond to local perturbations, so the two describe different aspects of the denoiser and can complement each other. The feature layer and sensitivity noise levels are chosen from 500 in-distribution calibration images only: geometrically transformed images serve as proxy shifts to rank layers, and the coefficient of variation of ID scores ranks noise levels; AUROC is computed on disjoint sets of 500 ID and 500 OOD test images.
Both scores are standardized on the in-distribution calibration set and then fused with a weight also fixed from in-distribution data, so the detector needs no training, no labels and no OOD samples. Reconstruction alone fails on pairs where its ordering inverts; sensitivity alone averages 0.856 AUROC on the 12 pairs, below the 0.897 of the fused score. On the 12 pairs, reconstruction alone averages 0.704 AUROC and sensitivity alone 0.856, against 0.897 for the deployed fusion; varying the probe count, retained eigenvalues, noise repetitions and probe scale over the reported ranges keeps average AUROC between 0.892 and 0.901.
Perspective
The demonstrated setting is a frozen diffusion model trained on in-distribution data, covering the 12 image pairs among CIFAR-10, CIFAR-100, SVHN and CelebA, plus near-OOD, the FFHQ-256 high-resolution benchmark, ImageNet-64 and additional OOD datasets such as LSUN, iSUN and Textures; calibration uses 500 ID images and testing uses disjoint sets of 500 ID and 500 OOD images. The authors list extending the same readout to foundation diffusion models with only lightweight ID calibration or adaptation as future work.
The method requires a diffusion model trained on the ID distribution, which can limit practical use when ID data are scarce or training a diffusion model is expensive. The sensitivity channel needs additional forward passes to estimate feature-Jacobian spectra; although far cheaper than power-iteration methods, it is still more expensive than a single reconstruction score. In the D4RL trajectory study, medium trajectories rank near chance as OOD against medium-replay ID data, which the authors attribute to the two resembling each other as sub-populations.
