Research map / Label-efficient and robust learning / OOD and anomaly detection
Diffusion-based Layer-wise Semantic Reconstruction for Unsupervised Out-of-Distribution 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 DLSR/LFDN, a DDIM-style latent diffusion network that reconstructs multi-layer EfficientNet-b4 pooled features after Gaussian distortion, using reconstruction error (MFsim, MSE, LR) as the unsupervised OOD score.
Problem
Unsupervised OOD detection trained only on unlabeled ID data; existing reconstruction-based generative methods trade off reconstruction power against compact ID representation, and pixel-level diffusion detectors are slow.
Approach
Perform diffusion-based denoising reconstruction on multi-layer, globally pooled and z-score normalized encoder features rather than pixels; score OOD by cosine similarity (MFsim), MSE, or likelihood regret between original and reconstructed features.
Key results
- On CIFAR-10 as ID (EfficientNet-b4), ours(+MFsim) reaches 99.34% average AUROC vs 97.44% for DICE.
- On CIFAR-100 as ID, ours(+MFsim) reaches 97.20% average AUROC vs 83.72% for DICE (about 13.5 points higher).
- Ours(+MFsim) runs at roughly 960 images/s on EfficientNet-b4, and the paper reports it is nearly 100x faster than pixel-level DDPM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use multi-layer semantic features rather than only the last encoder block (ablation) | DLSR with MFsim · CIFAR-10, CIFAR-100, and CelebA OOD benchmarks | AUROC: 99.93 -> 99.98; 79.24 -> 97.2; 93.84 -> 99.35 | Helped |
| replace the feature-level diffusion network with an autoencoder (ablation) | DLSR · CIFAR-10; OOD average | AUROC: 81.2 -> 99.7±0.1 | Hurt |
| reduce LFDN linear-layer dimension from 720 to 1440 (ablation) | DLSR with MSE · CIFAR-100; OOD average | AUROC: 83.35 -> 86.35 | Helped |
| reduce LFDN linear-layer dimension from 720 to 1440 (ablation) | DLSR with LR · CIFAR-100; OOD average | AUROC: 84.05 -> 89.17 | Helped |
| reduce LFDN linear-layer dimension from 720 to 1440 (ablation) | DLSR with MFsim · CIFAR-100; OOD average | AUROC: 96.43 -> 97.20 | Helped |
| increase LFDN residual blocks from 8 to 16 (ablation) | DLSR with MSE · CIFAR-100; OOD average | AUROC: 85.26 -> 86.35 | Helped |
| increase LFDN residual blocks from 8 to 16 (ablation) | DLSR with LR · CIFAR-100; OOD average | AUROC: 87.32 -> 89.17 | Helped |
| increase LFDN residual blocks from 8 to 16 (ablation) | DLSR with MFsim · CIFAR-100; OOD average | AUROC: 97.13 -> 97.20 | No clear effect |
| evaluate at a sampling time step near 3–10 rather than t=1 or very large t (ablation) | DLSR · CIFAR-10; six OOD datasets | AUROC | Mixed |
Surprising findings
- Pixel-level DDPM reconstruction error distributions for ID and OOD samples are very similar, whereas feature-level reconstruction separates them clearly.
- Reconstruction performance peaks at small sampling steps (t around 3-10); t=1 performs poorly because too little noise is added.
Limitations
- Performance depends on the quality of features from the pretrained encoder, as the authors note.
- MFsim performance drops on some OOD sets, e.g., CIFAR-10 vs CIFAR-100 (85.60 AUROC, below MSE's 87.54), and LR is slower than MSE/MFsim.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 CelebA ImageNet100 SVHN LSUN iSUN Textures Places365 iNaturalist MNIST FMNIST
Benchmarks: CIFAR-10 vs six OOD sets CIFAR-100 vs six OOD sets CelebA vs four OOD sets ImageNet100 vs four OOD sets CIFAR-10 vs CIFAR-100 DiffGuard CIFAR-10 vs CIFAR-100/TinyImageNet
Metrics: AUROC FPR95 images per second
Baselines: GLOW PixelCNN++ VAE DDPM AutoEncoder MSP EBO DICE ASH-S SimCLR+Mahalanobis SimCLR+KNN LMD
Builds on: DDPM DDIM Likelihood Regret Latent Diffusion Models
Closest papers by meaning
- Diffusion-based Semantic Outlier Generation via Nuisance Awareness for Out-of-Distribution Detection (AAAI 2025 · similarity 0.89)
- 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 0.88)
- The Best of Both Worlds: On the Dilemma of Out-of-distribution Detection (NeurIPS 2024 · similarity 0.87)
- BOOD: Boundary-based Out-Of-Distribution Data Generation (ICML 2025 · similarity 0.87)
- Image Background Serves as Good Proxy for Out-of-distribution Data (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.