Research map / Continual learning / Online and replay-based
One-for-More: Continual Diffusion Model 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 CDAD, a continual latent diffusion model for multi-class anomaly detection that uses gradient projection with an iterative SVD (iSVD) to prevent forgetting and an anomaly-masked network to condition the denoiser.
Problem
Diffusion-based anomaly detection suffers from catastrophic forgetting and faithfulness hallucination when trained continually on new anomaly-detection classes, and gradient-projection defenses are memory-prohibitive for diffusion U-Nets.
Approach
Projects new-task gradients orthogonal to the significant subspace of previous tasks, computes that subspace with an iterative SVD that splits features into blocks (memory roughly independent of image count), and uses an anomaly-masked conditional network with patch perturbation so the diffusion model reconstructs anomalous regions rather than overfitting to normal images.
Key results
- On MVTec continual settings, CDAD achieves best A-AUROC in all four settings (e.g., 96.4/96.6 in 14-1 with 1 step), with negative forgetting measure in that setting.
- iSVD with n=10 reduces memory for computing the significant representation of ten images from 157.3 GB to 16.7 GB (89.4% saving).
- On cross-dataset continual learning, CDAD outperforms the second-best method by 4.5/1.9 (MVTec-VisA) and 5.3/1.9 (VisA-MVTec) A-AUROC points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CDAD instead of the strongest reported baseline | continual anomaly detection · MVTec 3 × 5 with 5 steps | image A-AUROC: 84.8 -> 89.1 | Helped |
| Use CDAD instead of the strongest reported baseline | continual anomaly detection · VisA 8–1 × 4 with 4 steps | image A-AUROC: 79.8 -> 83.4 | Helped |
| Add the anomaly-masked network (AMN) (ablation) | CDAD · MVTec 10–1 × 5 with 5 steps | image A-AUROC: 94.3 -> 94.9 | Helped |
| Add the anomaly-masked network (AMN) (ablation) | CDAD · MVTec 10–1 × 5 with 5 steps | pixel A-AUROC: 94.0 -> 95.7 | Helped |
| Add the continual diffusion model (CDM) (ablation) | CDAD · MVTec 10–1 × 5 with 5 steps | image A-AUROC: 90.6 -> 94.9 | Helped |
| Add the continual diffusion model (CDM) (ablation) | CDAD · MVTec 10–1 × 5 with 5 steps | pixel A-AUROC: 93.5 -> 95.7 | Helped |
| Use iterative SVD instead of full SVD (ablation) | significant-representation computation · 10 images | memory: 157.3GB -> 16.7GB | Helped |
| Increase the iSVD split-block count (ablation) | iSVD · feature-matrix projection | projection performance | Hurt |
| Increase the k-rank approximation threshold γth (ablation) | CDAD · MVTec setting 2 | A-AUROC | Mixed |
| Use AMN as the condition mechanism (ablation) | CDAD · MVTec 10–1 × 5 with 5 steps | forgetting measure: 1.69 / 1.55 -> 1.01 / 0.74 | Helped |
Surprising findings
- Without the continual diffusion model (only AMN), forgetting is still lower than ControlNet, and the model sometimes improves on previous tasks after continual training.
- Larger iSVD split counts degrade projection quality, but the authors judge the information loss acceptable for preserving plasticity.
Limitations
- Continual learning does not project gradients fully into the orthogonal space, trading off plasticity for stability.
- iSVD loses information as the number of split blocks grows, degrading subspace quality.
Experimental setup
Datasets: MVTec AD VisA
Benchmarks: MVTec AD VisA
Metrics: A-AUROC FM AUROC
Baselines: UniAD UniAD + EWC UniAD + SI UniAD + MAS UniAD + LVT UCAD IUF DiAD ControlNet
Builds on: Stable Diffusion ControlNet gradient projection (GPM) UniAD IUF
Closest papers by meaning
- On Diffusion Modeling for Anomaly Detection (ICLR 2024 · similarity 0.89)
- Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.88)
- Unsupervised Continual Anomaly Detection with Contrastively-Learned Prompt (AAAI 2024 · similarity 0.87)
- AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · similarity 0.87)
- OmiAD: One-Step Adaptive Masked Diffusion Model for Multi-class Anomaly Detection via Adversarial Distillation (ICML 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.