SynSur: An end-to-end generative pipeline for synthetic industrial surface defect generation and detection

2026-01-26 · Source: cs.CV updates on arXiv.org · Field: Manufacturing & Industrial — Artificial Intelligence & Machine Learning, Smart Manufacturing & Industry 4.0, Quality Control & Standards · Depth: Expert, extended

Summary

SynSur is an end-to-end generative pipeline for synthetic industrial surface defect generation and detection, addressing the scarcity of labeled defect data. Developed by Fraunhofer IGD, it combines Vision-Language Model-based prompts, LoRA-adapted Flux.1-dev diffusion, mask-guided inpainting, and sample filtering with automatic label derivation via SAM 3. Evaluated on BSData (pitting defects on ball screw drives) and MSD (mobile phone screen scratches), the pipeline generates realistic defects. Experiments with YOLOv26, YOLOX, and LW-DETR show that synthetic-only training is inferior, but combining synthetic data with real data can preserve performance and yield modest gains, particularly for YOLO-style architectures. LoRA adaptation is crucial for domain-specific appearance matching, and cross-domain transfer requires careful adaptation.

Key takeaway

For AI Engineers developing industrial defect detection systems with limited real data, integrate synthetic data as an augmentation strategy rather than a full replacement. Focus on combining synthetic samples with your existing real datasets to achieve modest performance gains, especially with YOLO-style architectures. Ensure domain-specific adaptation, like LoRA fine-tuning and careful annotation quality control, for reliable results and cross-domain transferability.

Key insights

Synthetic data, when augmenting scarce real datasets, can improve industrial defect detection.

Principles

• LoRA adaptation is essential for domain-specific defect appearance matching.
• Synthetic data strengthens scarce real datasets rather than replacing them.

Method

The SynSur pipeline uses VLM-derived prompts, LoRA-adapted diffusion, mask-guided inpainting, DreamSim/CLIPScore filtering, and SAM 3 for automatic label derivation.

In practice

• Use LoRA for domain adaptation in diffusion models for specific defect types.
• Filter synthetic samples with DreamSim and CLIPScore to ensure quality and relevance.

Topics

• Industrial Defect Detection
• Synthetic Data Generation
• Diffusion Models
• LoRA
• Vision-Language Models
• BSData Dataset

Code references

• black-forest-labs/flux

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• SynSur: An end-to-end generative pipeline for synthetic industrial surface defect generation and detection — Synthetic data generation can augment scarce real datasets for industrial defect detection, improving model performance.
• XCT-SAM: Sequential Parameter-Efficient Domain Adaptation of SAM for Industrial XCT Defect Segmentation — Sequential parameter-efficient domain adaptation effectively bridges large domain gaps for specialized image segmentation.
• 3D-DefectBench: A Controlled Factorial Study of Vision-Language Model Evaluation Pipelines for Fine-Grained 3D Generation Defects — VLM-based 3D defect detection requires evaluating the entire pipeline, not just the model, for reliable automation.
• Hybrid Deep Learning for Traceability and Classification of Industrial Slate Tiles — A hybrid deep learning model combining feature matching and classification significantly improves industrial slate tile traceability and quality control.
• Balancing Real and Synthetic Data for CNN-based Masonry Crack Detection — Synthetic data can significantly reduce real data collection needs for CNN-based crack detection while maintaining or improving accuracy.
• Auditing Training Data in Domain-adapted LLMs: LoRA-MINT — LoRA-MINT uses perplexity and synthetic samples to audit training data membership in LoRA-fine-tuned LLMs.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.