Research map / Image generation / Autoregressive and latent generation

A Unified Latent Schrodinger Bridge Diffusion Model for Unsupervised Anomaly Detection and Localization

CVPR 2025 · Image generation › Autoregressive and latent generation · approach: Reconstruction-based diffusion image detection · 9 citations (OpenAlex)

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 LASB, a latent-space linear Schrödinger bridge diffusion model trained only on normal images that transforms anomalous latents into normal ones for unsupervised anomaly detection and localization on MVTec-AD and VisA.

Problem

Unsupervised anomaly detection and localization in industrial images, where reconstruction-based methods rely on auxiliary networks or specialized adaptations and standard Gaussian diffusion struggles to preserve structure during anomaly-to-normal reconstruction.

Approach

Replaces Gaussian diffusion with a linear Schrödinger bridge in a VQ-VAE latent space, using Dirac-delta boundary conditions so the bridge drives anomalous latents toward normal ones with semi-degradation that preserves structure; anomalies are localized by comparing multi-scale ImageNet features of input and reconstruction, following DiAD.

Key results

• On MVTec-AD, LASB reaches image-level AUROC/AP of 99.2%/99.3% and pixel-level AUROC/AP of 98.6%/78.2%, with pixel AP improved by 17.4 points over the second-best diffusion method in Table 2.
• On VisA (multi-class), LASB achieves 94.2% image AUROC and 98.2% pixel AUROC, outperforming DiAD (86.8% / 96.0%) and MambaAD in the reported comparison.
• LASB inference takes 0.74 s per image on a V100 at 10 NFEs, about 2x faster than DiAD and 1.6x faster than TransFusion, with up to 5x faster training than SB-based models.

Measured effects reported in the paper

Surprising findings

• Performance saturates at about 10 sampling function evaluations, so additional NFEs beyond 10 yield no further gains on either dataset.
• Standard pixel-space DDPM and LDM underperform strongly on localization (e.g., LDM pixel AUROC 85.1%), indicating that Gaussian diffusion is poorly suited to anomaly-free reconstruction.

Limitations

• Pixel-level APseg and F1max on MVTec-AD are weaker than some baselines in some settings, and the localization pipeline depends on DiAD-style multi-scale feature differencing with a pre-trained ImageNet model.
• Evaluated only on two industrial datasets (MVTec-AD and VisA); no medical or other domain evaluation.
• Several table entries are inconsistent across the paper (e.g., differing DRAEM and DiAD numbers between Table 1 and Table 3), and the key-result numbers are reported with mixed metrics.

Experimental setup

Datasets: MVTec-AD VisA

Benchmarks: MVTec-AD VisA

Metrics: AUROC AP F1max

Baselines: SimpleNet PatchCore DSR PaDiM CS-Flow CFLOW-AD OCR-GAN DRAEM RD4AD ADSPR DiffAD D3AD

Builds on: Schrödinger Bridge score-based generative models latent diffusion model DiAD DRAEM anomaly augmentation VQ-VAE

Closest papers by meaning

• Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.89)
• A Diffusion-Based Framework for Multi-Class Anomaly Detection (AAAI 2024 · similarity 0.88)
• AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · similarity 0.87)
• Dual-Interrelated Diffusion Model for Few-Shot Anomaly Image Generation (CVPR 2025 · similarity 0.86)
• Generating and Reweighting Dense Contrastive Patterns for Unsupervised Anomaly Detection (AAAI 2024 · similarity 0.86)
• DZAD: Diffusion-based Zero-shot Anomaly Detection (AAAI 2025 · similarity 0.86)