Research map / Image generation / Autoregressive and latent generation
BOOD: Boundary-based Out-Of-Distribution Data Generation
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 BOOD, which synthesizes image-level OOD data by perturbing ID features nearest the decision boundary in a CLIP-aligned latent space and decoding them with Stable Diffusion for OOD detection.
Problem
Automatically generating auxiliary outlier data for OOD detection that lies near class decision boundaries, which prior diffusion-based methods like DreamOOD cannot guarantee.
Approach
Identify ID features closest to the decision boundary by counting adversarial gradient-ascent steps needed to flip the prediction, then push those features past the boundary (plus c extra steps) to form OOD embeddings, which replace the class token embedding in a Stable Diffusion prompt to generate outlier images.
Key results
- On CIFAR-100, BOOD reduces average FPR95 from 40.31% (DreamOOD) to 10.67% and raises average AUROC from 90.15% to 97.42%.
- On ImageNet-100, BOOD achieves average FPR95 35.37% and AUROC 92.44%, versus DreamOOD's 38.76% and 92.02%.
- Ablation on CIFAR-100: removing boundary identification yields FPR95 99.61%, removing feature perturbation yields 44.26%, full BOOD yields 10.67%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BOOD instead of DreamOOD | OOD detection with CIFAR-100 as ID data · CIFAR-100 (average over 5 OOD datasets) | FPR95: 40.31 -> 10.67 | Helped |
| use BOOD instead of DreamOOD | OOD detection with CIFAR-100 as ID data · CIFAR-100 (average over 5 OOD datasets) | AUROC: 90.15 -> 97.42 | Helped |
| use BOOD instead of DreamOOD | OOD detection with ImageNet-100 as ID data · ImageNet-100 (average over 4 OOD datasets) | FPR95: 38.76 -> 35.37 | Helped |
| use BOOD instead of DreamOOD | OOD detection with ImageNet-100 as ID data · ImageNet-100 (average over 4 OOD datasets) | AUROC: 92.02 -> 92.44 | Helped |
| add boundary feature identification (ablation) | BOOD feature synthesis · CIFAR-100 (average) | FPR95: 44.26 -> 10.67 | Helped |
| add feature perturbation (ablation) | BOOD feature synthesis · CIFAR-100 (average) | FPR95: 99.61 -> 10.67 | Helped |
| replace Gaussian noise with BOOD's boundary-directed perturbation (ablation) | OOD feature synthesis · CIFAR-100 (average) | FPR95: 18.99 -> 10.67 | Helped |
| replace centroid displacement with BOOD's boundary-directed perturbation (ablation) | OOD feature synthesis · CIFAR-100 (average) | FPR95: 40.51 -> 10.67 | Helped |
| use BOOD instead of DreamOOD | OOD detection with ImageNet-100 as ID data · NearOOD average (NINCO and SSB-hard) | FPR95: 67.19 -> 64.93 | Helped |
| use ResNet-18 instead of ResNet-34 (ablation) | BOOD · CIFAR-100 | FPR95: 10.67 -> 10.83 | No clear effect |
| use ResNet-50 instead of ResNet-34 (ablation) | BOOD · CIFAR-100 | FPR95: 10.67 -> 11.23 | No clear effect |
Surprising findings
- Adding Gaussian noise or displacing features from class centroids performs worse than boundary-based perturbation (FPR95 18.99 and 40.51 vs 10.67 on CIFAR-100).
- Removing boundary identification alone (decoding selected ID features directly) yields near-chance OOD detection (AUROC 7.83 on CIFAR-100).
Limitations
- Performance may degrade for datasets with large domain discrepancy from the diffusion model's training distribution.
- Classifier errors on unseen outlier features can mislead the decision-boundary crossing check, producing low-quality OOD features.
- Hyperparameters (step size, c, K, r, beta) are difficult to tune; the number of extra steps c requires manual selection.
Experimental setup
Datasets: CIFAR-100 ImageNet-100 SVHN Places365 Textures LSUN iSUN iNaturalist SUN Places NINCO SSB-hard
Benchmarks: CIFAR-100 OOD benchmark (SVHN, Places365, LSUN, iSUN, Textures) ImageNet-100 OOD benchmark (iNaturalist, Places, SUN, Textures) NINCO SSB-hard
Metrics: FPR95 AUROC ID ACC
Baselines: MSP ODIN Mahalanobis Energy GODIN KNN ViM ReAct DICE GAN VOS NPOS
Builds on: DreamOOD VOS CLIP Stable Diffusion
Closest papers by meaning
- Fast Decision Boundary based Out-of-Distribution Detector (ICML 2024 · similarity 0.89)
- Diffusion-based Semantic Outlier Generation via Nuisance Awareness for Out-of-Distribution Detection (AAAI 2025 · similarity 0.89)
- Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.88)
- GOOD: Training-Free Guided Diffusion Sampling for Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.88)
- Out-of-Distribution Detection with a Single Unconditional Diffusion Model (NeurIPS 2024 · similarity 0.88)
- Extremely Simple Multimodal Outlier Synthesis for Out-of-Distribution Detection and Segmentation (NeurIPS 2025 · similarity 0.88)
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.