Research map / Label-efficient and robust learning / OOD and anomaly detection
DeCoFlow: Structural Decomposition of Normalizing Flows for Continual Anomaly Detection
hun im, Jungi Lee, Subeen Cha, Pilsung Kang
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 |
|---|---|---|---|
| decompose coupling subnets into frozen base + LoRA instead of other architectures with frozen+LoRA (ablation) | normalizing flow · MVTec-AD | I-AUC: 62.03-68.15 -> 98.47 | Improved |
| decompose at feature level instead of coupling level (ablation) | DeCoFlow NF · MVTec-AD | I-AUC: 98.47 -> 50.00 | Hurt |
| remove ACL (ablation) | DeCoFlow · MVTec-AD | I-AUC: -10.17 %p when removed | Improved |
| add Tail-Aware Loss (ablation) | DeCoFlow · MVTec-AD | I-AUC: +3.17 %p (full model vs without TAL) | Improved |
| add Task-Specific Alignment (ablation) | DeCoFlow · MVTec-AD | I-AUC: +0.68 %p (full vs without TSA) | Improved |
| use LoRA instead of full per-task subnet (ablation) | DeCoFlow · MVTec-AD | I-AUC: -0.04 %p when LoRA excluded; doubles per-task params | No clear effect |
| apply TAL when base is frozen vs trainable (ablation) | normalizing flow · MVTec-AD | I-AUC: Frozen +13.04 %p; Trainable -1.22 %p | Mixed |
| add TSA under trainable base (ablation) | normalizing flow · MVTec-AD | I-AUC: -16.34 %p (trainable); -1.19 %p (frozen) | Hurt |
| replace ACL with additional DCL blocks (8DCL+0ACL) (ablation) | DeCoFlow · MVTec-AD | I-AUC/performance: -9.01 %p | Hurt |
| increase LoRA rank from 16 to 128 (ablation) | DeCoFlow · MVTec-AD | I-AUC / P-AP: 0.02 / 0.09 %p change | No clear effect |
| update shared base or use high-rank adapters (ablation) | DeCoFlow · MVTec-AD | I-AUC: 4.7-25.0x parameter cost, no significant gains | No clear effect |
| force 5% misrouting of tasks (ablation) | DeCoFlow · MVTec-AD | P-AP: 56.93 -> 30.20 | Hurt |
Closest papers by meaning
- InCLAD: A Continual Learning Benchmark for Industrial Visual Anomaly Detection (NeurIPS 2026 · similarity 0.86)
- SANFlow: Semantic-Aware Normalizing Flow for Anomaly Detection (NeurIPS 2023 · similarity 0.86)
- Rethinking Continual Anomaly Detection on the Edge: Benchmarking Under Realistic Industrial Conditions (ECCV 2026 · similarity 0.85)
- Unsupervised Continual Anomaly Detection with Contrastively-Learned Prompt (AAAI 2024 · similarity 0.83)
- One-for-More: Continual Diffusion Model for Anomaly Detection (CVPR 2025 · similarity 0.83)
- Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection (CVPR 2025 · similarity 0.83)
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.
