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

Industrial anomaly segmentation relies heavily on pixel-level annotations, yet real-world anomalies are often scarce, diverse, and costly to label. Segmentation-oriented industrial anomaly synthesis (SIAS) has emerged as a promising alternative; however, existing methods struggle to balance sampling efficiency and generation quality. Moreover, most approaches treat all spatial regions uniformly, overlooking the distinct statistical differences between anomaly and background areas. This uniform treatment hinders the synthesis of controllable, structure-specific anomalies tailored for segmentation tasks. In this paper, we propose FAST, a foreground-aware diffusion framework featuring two novel modules: the Anomaly-Informed Accelerated Sampling (AIAS) and the Foreground-Aware Reconstruction Module (FARM). AIAS is a training-free sampling algorithm specifically designed for segmentation-oriented industrial anomaly synthesis, which accelerates the reverse process through coarse-to-fine aggregation and enables the synthesis of state-of-the-art segmentation-oriented anomalies in as few as 10 steps. Meanwhile, FARM adaptively adjusts the anomaly-aware noise within the masked foreground regions at each sampling step, preserving localized anomaly signals throughout the denoising trajectory. Extensive experiments on multiple industrial benchmarks demonstrate that FAST consistently outperforms existing anomaly synthesis methods in downstream segmentation tasks. We release the code at: https://github.com/Chhro123/fast-foreground-aware-anomaly-synthesis.

💡 Research Summary

The paper tackles a critical bottleneck in industrial visual inspection: the scarcity of pixel‑accurate anomaly (defect) annotations. While synthetic anomaly generation—referred to as Segmentation‑oriented Industrial Anomaly Synthesis (SIAS)—has emerged as a promising remedy, existing approaches suffer from three intertwined problems. First, they provide limited controllability over the shape, location, and extent of generated defects, especially GAN‑based one‑shot generators. Second, training‑free manipulations (e.g., patch replacement, texture blending) can produce visible anomalies but lack the structural complexity and contextual consistency of real industrial defects, which hampers downstream segmentation performance. Third, diffusion‑based methods, although powerful, treat every pixel uniformly during both forward and reverse processes, ignoring the distinct statistical properties of anomaly versus background regions. This uniform treatment leads to anomaly signals being gradually diluted across the denoising trajectory, and the sampling process typically requires hundreds to thousands of steps, making it impractical for real‑time production line deployment.

FAST (Foreground‑Aware Diffusion with Accelerated Sampling Trajectory) addresses these shortcomings with two complementary modules: Anomaly‑Informed Accelerated Sampling (AIAS) and the Foreground‑Aware Reconstruction Module (FARM).

AIAS – training‑free, coarse‑to‑fine acceleration
Standard DDPMs define a discrete‑time Markov chain with a fixed variance schedule. The reverse transition from step t to t‑1 can be expressed analytically as a linear‑Gaussian function of the clean image x₀ and the current noisy latent xₜ. The authors observe that, when the model’s predicted clean image ˆx₀ is accurate (which is encouraged by the DDPM loss) and varies slowly over a short temporal window, it can be treated as constant within a segment of timesteps