Research map / Image generation / Fast diffusion sampling
On Diffusion Modeling for Anomaly Detection
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 Diffusion Time Estimation (DTE), which estimates the posterior over diffusion time of noisy inputs as an anomaly score, giving a kNN-equivalent non-parametric variant and fast MLP/ResNet parametric variants, and evaluates on ADBench against classical, deep, and diffusion baselines.
Problem
Diffusion-based anomaly detection (e.g., DDPM reconstruction error) is accurate but computationally expensive, and has been explored mainly for image data rather than general tabular, image-embedding and NLP-embedding anomaly detection in unsupervised and semi-supervised settings.
Approach
Instead of simulating the full DDPM reverse chain to reconstruct inputs, model the posterior distribution over diffusion time (noise variance) of a noisy input, which has an inverse Gamma form; its mode/mean serves as the anomaly score. A non-parametric estimator reduces to kNN ranking, and a neural network predicts the distribution parametrically for fast inference.
Key results
- DTE (non-parametric and parametric) is competitive with kNN and classical methods across 57 ADBench datasets in semi-supervised and unsupervised settings, and outperforms DDPM and prior deep baselines overall.
- Parametric DTE achieves orders-of-magnitude faster inference time than DDPM and other methods, with the shortest inference among all methods considered.
- Using pre-trained embeddings instead of raw pixels improves diffusion-based methods on image anomaly detection; e.g., on CIFAR10 with VicReg pre-training DTE-NP reaches 80.92 AUC vs 50.50 with no pre-training for the no-pretraining baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use VicReg pre-training instead of no pre-training (ablation) | DTE-NP · CIFAR10, semi-supervised | AUC ROC: 51.53(14.81) -> 80.92(10.81) | Helped |
| Use VicReg pre-training instead of no pre-training (ablation) | DTE-C · CIFAR10, semi-supervised | AUC ROC: 50.25(3.34) -> 63.36(11.92) | Helped |
| Use VicReg pre-training instead of no pre-training (ablation) | DDPM · CIFAR10, semi-supervised | AUC ROC: 50.50(7.67) -> 54.22(10.26) | No clear effect |
| Use VicReg pre-training instead of no pre-training (ablation) | kNN · CIFAR10, semi-supervised | AUC ROC: 51.64(14.90) -> 79.01(11.53) | Helped |
| Use supervised ImageNet pre-training instead of no pre-training (ablation) | DTE-NP · MNIST, semi-supervised | AUC ROC: 81.94(16.46) -> 78.07(12.48) | Hurt |
| Use supervised ImageNet pre-training instead of no pre-training (ablation) | kNN · MNIST, semi-supervised | AUC ROC: 84.14(15.60) -> 76.86(11.54) | Hurt |
| Use image embeddings rather than raw pixels (ablation) | Diffusion-based methods · Image datasets | AUC ROC | Helped |
| Use pre-training rather than no pre-training (ablation) | Image anomaly detection methods · VisA, CIFAR10, and MNIST | AUC ROC | Mixed |
| Increase the DDPM reconstruction timestep (ablation) | DDPM · ADBench, semi-supervised | AUC ROC | Hurt |
| Set DDPM reconstruction timestep close to 50% of the maximum (ablation) | DDPM · ADBench, unsupervised | AUC ROC | Helped |
| Use fewer bins in the categorical model (ablation) | DTE-C · ADBench | AUC ROC | Helped |
| Increase the maximum diffusion timestep from very low values toward roughly T=250 (ablation) | DTE · ADBench | AUC ROC | Helped |
Surprising findings
- The non-parametric DTE yields the same anomaly ranking as kNN, so diffusion-based scoring reduces to a kNN-like method.
- Classical kNN remains a top-performing method, outperforming DTE in some settings such as the semi-supervised ADBench comparison.
Limitations
- Interpretability of DTE is weaker than DDPM's denoising-based explanations.
- Only point anomalies are addressed; group and contextual anomalies are left for future work.
- Diffusion methods lag specialized methods on VisA, and kNN can outperform DTE in some settings.
Experimental setup
Datasets: ADBench VisA CIFAR-10 MNIST SVHN FashionMNIST MNIST-C
Benchmarks: ADBench VisA
Metrics: AUC ROC AUC PR F1 score inference time
Baselines: kNN DDPM DeepSVDD DAGMM IForest ECOD COPOD DIF SLAD planar normalizing flows VAE GAN
Builds on: DDPM kNN TabDDPM ADBench
Closest papers by meaning
- One-for-More: Continual Diffusion Model for Anomaly Detection (CVPR 2025 · similarity 0.89)
- OmiAD: One-Step Adaptive Masked Diffusion Model for Multi-class Anomaly Detection via Adversarial Distillation (ICML 2025 · similarity 0.88)
- Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.88)
- Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection (CVPR 2025 · similarity 0.87)
- Scalable, Explainable and Provably Robust Anomaly Detection with One-Step Flow Matching (NeurIPS 2025 · similarity 0.86)
- A Diffusion-Based Framework for Multi-Class Anomaly Detection (AAAI 2024 · similarity 0.86)
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.