Research map / Image generation / Autoregressive and latent generation
InvAD: Inversion-based Reconstruction-Free Anomaly Detection with Diffusion Models
Shunsuke Sakai, Xiangteng He, Chunzhi Gu, Leonid Sigal, Tatsuhito Hasegawa
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-step inversion (instead of single-step) (ablation) | InvAD with FDM · MVTecAD | mAD: 71.0 -> 83.7 | Improved |
| add inversion to feature-space diffusion (ablation) | InvAD · MVTecAD | mAD: 57.3 -> 83.7 | Improved |
| add feature-space diffusion to single-step inversion (ablation) | InvAD · MVTecAD | mAD: 44.9 -> 71.0 | Improved |
| add NLL+Diff scoring instead of NLL alone at S=1000 (ablation) | InvAD · MVTecAD | image AU-ROC: 89.1 -> 95.4 | Improved |
| add Diff term to NLL at S=3 (ablation) | InvAD scoring · MVTecAD | image AU-ROC: 96.1 -> 99.0 | Improved |
| add NLL term to Diff at S=1000 (ablation) | InvAD scoring · MVTecAD | image AU-ROC: 82.0 -> 95.4 | Improved |
| use inversion instead of reconstruction at S=3 (ablation) | diffusion AD · MVTecAD | image AU-ROC: 64.9 -> 99.0 | Improved |
| increase diffusion steps S from 3 to 1000 (ablation) | InvAD · MVTecAD | image AU-ROC: 99.0 -> 95.4 | Hurt |
| plug InvAD into DiAD | DiAD · MVTecAD | image AU-ROC / FPS: 97.2 -> 98.2; 0.1 -> 88.1 | Improved |
| plug InvAD into MDM | MDM · MVTecAD | image AU-ROC / FPS: 91.9 -> 98.2; 2.2 -> 63 | Improved |
| use EfficientNet-B4 instead of ViT-B feature extractor (ablation) | InvAD · MVTecAD | image AU-ROC: 94.2 -> 99.0 | Improved |
| replace DiT-gigant with MLP diffusion network (ablation) | InvAD · MVTecAD | image AU-ROC: 99.0 -> 97.2 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- OmiAD — OmiAD: One-Step Adaptive Masked Diffusion Model for Multi-class Anomaly Detection via Adversarial Distillation (ICML 2025) · better on 2 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.90)
- A Diffusion-Based Framework for Multi-Class Anomaly Detection (AAAI 2024 · similarity 0.89)
- DZAD: Diffusion-based Zero-shot Anomaly Detection (AAAI 2025 · similarity 0.88)
- Adapt and Diffuse: Sample-adaptive Reconstruction via Latent Diffusion Models (ICML 2024 · similarity 0.87)
- InvFusion: Bridging Supervised and Zero-shot Diffusion for Inverse Problems (NeurIPS 2025 · similarity 0.87)
- On Diffusion Modeling for Anomaly Detection (ICLR 2024 · 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.
