" A Synthetic Dataset for Extreme Few-Shot Industrial Defect Inspection"

"This repository provides the official code implementation and experimental frameworks for lightweight dual-branch diffusion for few-shot industrial defect inspection. Designed specifically for resource-constrained edge devices (under 8GB VRAM), the codebase enables high-fidelity defect synthesis and downstream instance segmentation under extreme 10-shot constraints. The project is structured into two primary pipelines: a generative framework featuring efficient diffusion training and inference scripts, and a comprehensive downstream evaluation suite. The latter provides a complete generation-to-deployment workflow, including an interactive Human-In-The-Loop (HITL) quality screening tool, YOLOv8s-seg training scripts integrated with a Real-Data Anchoring (RDA) strategy to mitigate representation drift, and automated pipelines for calculating rigorous academic metrics (e.g., FID, KID, LPIPS, Pixel-AUROC). Together, these components allow researchers to fully reproduce our experiments across both standardized benchmarks and proprietary industrial datasets."