Research map / Image generation / Fast diffusion sampling

FAST: Foreground‑aware Diffusion with Accelerated Sampling Trajectory for Segmentation‑oriented Anomaly Synthesis

NeurIPS 2025 · Image generation › Fast diffusion sampling · approach: Distillation for few-step generation · 0 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 FAST, a training-free coarse-to-fine accelerated sampler (AIAS) plus a foreground-aware reconstruction module (FARM) for diffusion-based segmentation-oriented industrial anomaly synthesis, reaching strong downstream mIoU/Acc on MVTec-AD and BTAD with as few as 10 steps.

Problem

Industrial anomaly segmentation needs pixel-level labeled anomalies, which are scarce and costly; existing anomaly synthesis methods lack controllability, ignore segmentation-relevant structure, treat foreground and background uniformly, and are slow to sample.

Approach

AIAS aggregates many DDPM reverse steps into a few closed-form affine-Gaussian coarse-to-fine updates (training-free), while FARM reconstructs pseudo-clean anomaly content and re-injects anomaly-aware noise only in masked foreground regions at each step.

Key results

• On MVTec-AD with Segformer, FAST reaches average mIoU 76.72% and Acc 83.97%, versus 74.75% Acc for the strongest prior (DRAEM) as stated in the text.
• AIAS matches full 1000-step DDPM performance using only about 10 sampling steps and reaches near-optimal results at 50 steps.
• Adding FARM raises average MVTec mIoU from 65.33 to 76.72 and Acc from 71.24 to 83.97 in the ablation.

Measured effects reported in the paper

Surprising findings

• Fewer sampling steps (AIAS at 50 steps) outperform DDPM with 1000 steps on downstream segmentation, which the authors attribute to the evaluation objective favoring anomaly localization over visual fidelity.
• Table-reported FAST mIoU for MVTec/Segformer is 76.72 with DRAEM's Acc cited as 74.75, and the text inconsistently reports some numbers, indicating reporting ambiguity.

Limitations

• Very small final step counts (t=1 or 2) may leave residual artifacts, requiring a fine refinement stage.
• Mask generation relies on manual screening of synthesized masks, limiting automation.
• Evaluated only on MVTec-AD and BTAD with lightweight segmentation backbones.

Experimental setup

Datasets: MVTec-AD BTAD

Benchmarks: MVTec-AD BTAD

Metrics: mIoU pixel accuracy

Baselines: CutPaste DRAEM GLASS DFMGAN RealNet AnomalyDiffusion DDPM DDIM PLMS

Builds on: DDPM DDIM Anomaly Diffusion latent diffusion model

Closest papers by meaning

• AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · similarity 0.86)
• RealNet: A Feature Selection Network with Realistic Synthetic Anomaly for Anomaly Detection (CVPR 2024 · similarity 0.86)
• Dual-Interrelated Diffusion Model for Few-Shot Anomaly Image Generation (CVPR 2025 · similarity 0.86)
• Self-Supervised Likelihood Estimation with Energy Guidance for Anomaly Segmentation in Urban Scenes (AAAI 2024 · similarity 0.85)
• OmiAD: One-Step Adaptive Masked Diffusion Model for Multi-class Anomaly Detection via Adversarial Distillation (ICML 2025 · similarity 0.85)
• Multimodal Industrial Anomaly Detection by Crossmodal Feature Mapping (CVPR 2024 · similarity 0.85)