AI-powered visual inspection in industrial manufacturing suffers from the scarcity and subtleness of defect samples. Consequently, anomaly generation has become crucial for building robust inspection systems, yet it remains challenging when dealing wi...
AI-powered visual inspection in industrial manufacturing suffers from the scarcity and subtleness of defect samples. Consequently, anomaly generation has become crucial for building robust inspection systems, yet it remains challenging when dealing with fine-grained anomalies such as tiny, slender defects, which increasingly emerge as manufacturing processes continue to miniaturize. Despite recent progress in diffusion-based generators, it remains difficult to synthesize or preserve such fine structures, as standard latent diffusion models often sacrifice high-frequency spatial details during the encoding process.
In this work, we present a novel diffusion operator, Anomaly-generating Latent Operator with Hierarchical Attention (ALOHA), designed to preserve such fine-grained anomalies. Our key insight is to leverage Cross-scale Fusion using Receptive-Field–augmented Attention (RFAtten) to learn richly fine-grained anomaly information across all features coming from the hierarchical U-Net denoiser. This mechanism effectively bridges the information gap between shallow and deep layers, preventing the dilution of minute structural cues during the denoising steps. By dynamically calibrating the receptive fields, RFAtten enables the model to simultaneously capture fine-grained morphological structures while maintaining global semantic consistency. ALOHA integrates a ControlNet branch for mask-guided spatial conditioning to enable the generation of accurately-matched anomalous image-mask pairs and to achieve precise, mask-guided spatial control.
This synergistic architecture ensures that the synthesized defects not only blend seamlessly with the background texture but also strictly adhere to the morphological constraints of the input masks. Extensive experiments on the MVTec AD dataset demonstrate that ALOHA generates highly authentic and structurally coherent anomalies, outperforming state-of-the-art methods and significantly improving downstream anomaly inspection performance.