Research map / Image generation / Autoregressive and latent generation
Diffusion-based Semantic Outlier Generation via Nuisance Awareness for 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 SONA, which generates semantic-discrepant but nuisance-preserving outliers by guiding a Stable Diffusion model with ID images via region-specific (semantic/nuisance) guidance, and trains a ResNet-18 OOD detector with outlier-exposure and mutual-information losses.
Problem
Synthetic outlier generation for OOD detection, especially near-OOD, produces outliers too far from the in-distribution data and fails to capture subtle semantic distinctions; diffusion-based methods are sensitive to generation targets/prompts.
Approach
Starts from pixel-space ID images, partially noises them to an early stop timestep, and denoises with SONA guidance that removes ID semantics in semantic regions, preserves ID nuisance regions, and injects OOD-label semantics in the non-overlapping semantic region, then trains the classifier with CE, outlier-exposure, and a contrastive mutual-information upper-bound loss.
Key results
- On ImageNet-200, SONA reaches 88.38% average near-OOD AUROC and 95.87% far-OOD AUROC, versus 80.62% for DreamOOD on near-OOD.
- On ImageNet-1k, SONA achieves 85.32% near-OOD and 94.82% far-OOD average AUROC, outperforming DreamOOD (77.03 / 92.50).
- Ablation shows removing nuisance-preservation term drops near-OOD AUROC from 88.38 to 84.56 (with ID-removal and OOD-semantic terms only).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train the detector with SONA-generated outliers rather than prior methods | ResNet-18 OOD detector · ImageNet-200 Near-OOD (Avg) | AUROC: 84.84 -> 88.38±0.18 | Helped |
| Train the detector with SONA-generated outliers rather than prior methods | ResNet-18 OOD detector · ImageNet-200 Far-OOD (Avg) | AUROC: 94.49 -> 95.87±0.20 | Helped |
| Add ΔID to global guidance (ablation) | SONA outlier generation · ImageNet-200 Near-OOD | AUROC: 82.85 -> 83.74 | Helped |
| Add filtered ΔOOD to ΔID guidance (ablation) | SONA outlier generation · ImageNet-200 Near-OOD | AUROC: 83.74 -> 84.56 | Helped |
| Use unfiltered ΔOOD instead of nuisance-filtered ΔOOD (ablation) | SONA outlier generation · ImageNet-200 Near-OOD | AUROC: 84.56 -> 82.57 | Hurt |
| Add ΔN to ΔID and filtered ΔOOD (ablation) | SONA outlier generation · ImageNet-200 Near-OOD | AUROC: 84.56 -> 88.38 | Helped |
| Add LOE to LCE training (ablation) | ResNet-18 OOD detector · ImageNet-200 Near-OOD | AUROC: 82.5 -> 87.19 | Helped |
| Add LMI to LCE and LOE training (ablation) | ResNet-18 OOD detector · ImageNet-200 Near-OOD | AUROC: 87.19 -> 88.38 | Helped |
| Vary region masking threshold λ and guidance scale s over the tested ranges (ablation) | SONA guidance · ImageNet-200 | AUROC | No clear effect |
| Randomly select stop timestep T̃ rather than use a fixed timestep (ablation) | SONA guidance · ImageNet-200 | AUROC | Helped |
Surprising findings
- Generated outliers from ImageNet-800 prompts give similar results to LAION prompts, while DreamOOD is highly sensitive to prompt choice.
- Real-world outlier methods such as OE are outperformed by SONA on ImageNet-200 near-OOD without using any extra real outlier data.
Limitations
- Relies on a pretrained Stable Diffusion model and synthetic outlier generation cost (about 8.1 hours for 200K images).
- Only the reported hyperparameters are tested; the method's sensitivity to other diffusion backbones is not studied.
- Fine-grained results are presented only briefly and the authors note more precise semantic guidance may further help.
Experimental setup
Datasets: ImageNet-200 ImageNet-800 (OOD label prompts) CIFAR-100 ImageNet-1k NABirds
Benchmarks: SSB-hard NINCO iNaturalist Textures OpenImage-O TIN MNIST SVHN Place365 ImageNet-1k near/far OOD ImageNet-200 full-spectrum NABirds fine-grained
Metrics: AUROC ID accuracy
Baselines: OpenMax MSP ODIN EBO OpenGAN ReAct KNN DICE OE MCD UDG MixOE
Builds on: Classifier-free diffusion guidance Outlier Exposure DreamOOD Stable Diffusion Energy-based OOD detection
Closest papers by meaning
- Diffusion-based Layer-wise Semantic Reconstruction for Unsupervised Out-of-Distribution Detection (NeurIPS 2024 · similarity 0.89)
- BOOD: Boundary-based Out-Of-Distribution Data Generation (ICML 2025 · similarity 0.89)
- GOOD: Training-Free Guided Diffusion Sampling for Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.89)
- Rethinking the Evaluation of Out-of-Distribution Detection: A Sorites Paradox (NeurIPS 2024 · 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.