Research map / Image generation / Autoregressive and latent generation
Out-of-Distribution Detection with a Single Unconditional Diffusion Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DiffPath, which uses one unconditional diffusion model (ImageNet or CelebA) and scores OOD samples by the rate-of-change and curvature of their DDIM-inverted diffusion paths, achieving competitive AUROC across pairwise OOD tasks with ~10 NFEs.
Problem
Unsupervised OOD detection traditionally requires training a separate generative model per inlier dataset; the paper asks whether a single diffusion model can perform OOD detection across diverse inlier/outlier pairs.
Approach
Measure the norm of the score (first-order, rate-of-change) and of its time derivative (curvature) along the DDIM-inverted probability flow ODE path mapping samples to standard normal, fit a density (KDE or GMM) on inlier statistics, and use it as the OOD score; a 6-D sign-preserving variant resolves symmetric-path failures such as negative CIFAR10.
Key results
- DiffPath-6D-CelebA achieves an average AUROC of 0.931 across the 12 pairwise tasks with 10 NFEs, comparable to MSMA (0.945) and above the other baselines.
- DiffPath-6D-ImageNet achieves an average AUROC of 0.850 with 10 NFEs, competitive with LMD (0.865) which needs 10^4 NFEs.
- DiffPath-1D fails (AUROC 0.500) to separate CIFAR10 from negative CIFAR10, while DiffPath-6D reaches 0.994.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiffPath-6D with a CelebA-trained model instead of MSMA | OOD detection across the 12 pairwise tasks · 12 pairwise OOD tasks average | AUROC: 0.945 -> 0.931 | Hurt |
| Use DiffPath-6D with an ImageNet-trained model instead of LMD | OOD detection across the 12 pairwise tasks · 12 pairwise OOD tasks average | AUROC: 0.865 -> 0.850 | Hurt |
| Replace the 1D statistic with the 6D statistic (ablation) | DiffPath with an ImageNet-trained model · 12 pairwise OOD tasks average | AUROC: 0.806 -> 0.850 | Helped |
| Replace the 1D statistic with the 6D statistic (ablation) | DiffPath with a CelebA-trained model · 12 pairwise OOD tasks average | AUROC: 0.822 -> 0.931 | Helped |
| Train the diffusion model on CelebA rather than CIFAR10 (ablation) | DiffPath-6D · 12 pairwise OOD tasks average | AUROC: 0.853 -> 0.931 | Helped |
| Train the diffusion model on CelebA rather than SVHN (ablation) | DiffPath-6D · 12 pairwise OOD tasks average | AUROC: 0.838 -> 0.931 | Helped |
| Train the diffusion model instead of using random initialization (ablation) | DiffPath-6D · 12 pairwise OOD tasks average | AUROC: 0.501 -> 0.931 | Helped |
| Vary the DDIM integration steps from 10 to 5, 25, or 50 (ablation) | DiffPath-6D-CelebA · 12 pairwise OOD tasks average | AUROC: 0.929; 0.931; 0.919; 0.910 | No clear effect |
| Use inconsistent resizing instead of consistent resizing (ablation) | DiffPath-6D-ImageNet · 12 pairwise OOD tasks average | AUROC: 0.850 -> 0.942 | Mixed |
| Use DiffPath-6D rather than DiffPath-1D (ablation) | CIFAR10 vs negative CIFAR10 detection · CIFAR10 vs negative CIFAR10 | AUROC: 0.500 -> 0.994 | Helped |
| Use DiffPath-6D for near-OOD detection | Near-OOD detection · CIFAR10 near-OOD average and TinyImageNet near-OOD average | AUROC: 0.607; 0.845 | Mixed |
Surprising findings
- A randomly initialized diffusion model gives average AUROC 0.501, but a single model trained on ImageNet transfers to distributions it never saw during training.
- Naive resizing of images of different original resolutions inflates AUROC (e.g., CelebA vs CIFAR10 near-perfect), so the authors introduced a downsample-then-upsample resizing protocol.
Limitations
- Near-OOD performance is weak (CIFAR10 vs CIFAR100 AUROC about 0.59), and on OpenOOD CIFAR10 the method scores 0.607 versus 0.907 for the best discriminative baseline.
- Only the simple finite-difference time derivative is used rather than the full curvature.
- Evaluated only on natural-image DMs (CelebA, ImageNet), not on specialized domains such as medical images.
Experimental setup
Datasets: CIFAR10 SVHN CelebA ImageNet CIFAR100 Textures
Benchmarks: CIFAR10 vs SVHN CIFAR10 vs CelebA CIFAR10 vs CIFAR100 CIFAR10 vs Textures SVHN vs CIFAR10 SVHN vs CelebA SVHN vs CIFAR100 SVHN vs Textures CelebA vs CIFAR10 CelebA vs SVHN CelebA vs CIFAR100 CelebA vs Textures
Metrics: AUROC NFE
Baselines: IGEBM VAEBM Improved CD DoS WAIC Typicality Test Likelihood Ratio Input Complexity NLL DDPM-OOD LMD MSMA
Builds on: DDIM Improved DDPM MSMA Likelihood Regret
Closest papers by meaning
- Diffusion-based Semantic Outlier Generation via Nuisance Awareness for Out-of-Distribution Detection (AAAI 2025 · similarity 0.88)
- BOOD: Boundary-based Out-Of-Distribution Data Generation (ICML 2025 · similarity 0.88)
- GOOD: Training-Free Guided Diffusion Sampling for Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.88)
- Mysteries of the Deep: Role of Intermediate Representations in Out of Distribution Detection (NeurIPS 2025 · similarity 0.87)
- Out-Of-Distribution Detection with Diversification (Provably) (NeurIPS 2024 · similarity 0.87)
- Splitting & Integrating: Out-of-Distribution Detection via Adversarial Gradient Attribution (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.