What is it about?
AI systems that inspect products need examples of defects, but real faulty products can be rare. This makes it difficult to collect enough images to teach these systems what to look for. AnomalyBrush creates realistic images of defective products using a small number of real examples. It combines a photo of a defect-free product, a marked area showing where a defect should appear, and a short description of the defect. One part of the system draws useful details from real defects, while another adds a similar defect to the new image. Additional checks help the generated defect match the intended shape, location, and type while preserving the product's identity. The image generator does not need additional training for this process. Tests on two industrial inspection benchmarks, MVTec AD and Real-IAD, showed improvements in image quality and in how well generated defects matched their location labels. The generated images also helped an inspection model locate defects in real images more effectively.
Featured Image
Photo by Sam Moghadam on Unsplash
Why is it important?
Reliable quality inspection depends on recognising defects that may occur only rarely. Collecting enough real examples, and marking exactly where each defect appears, can be difficult and time-consuming. AnomalyBrush addresses both challenges by creating realistic defect images with matching location labels. This matters because a synthetic image can look convincing while still providing misleading information about the defect's exact position or shape. Our method focuses on keeping these details consistent, including for small defects and thin cracks. By making better use of limited real examples, this research offers a way to expand training data for industrial inspection. The results suggest that improving the quality and precision of synthetic data can help AI systems locate real defects more effectively.
Perspectives
For me, an important message of this work is that a convincing synthetic image is only part of the goal. If we want generated images to help train inspection systems, the labels must also accurately describe where the defects are. AnomalyBrush brings these two requirements together. I see it as a step toward making generative AI more useful for practical inspection tasks, especially when real defect examples are scarce.
Gaoxiang Li
Beijing Normal University
Read the Original
This page is a summary of: AnomalyBrush: Highly-Aligned Anomaly Image Generation with a Dual-Branch Diffusion Model, November 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3767308.3836539.
You can read the full text:
Contributors
The following have contributed to this page
