Training-Free Defect Generation

lastmod 2026-02-13created 2026-02-131 edits

Generates defect images without additional training using Stable Diffusion V2.1 and ControlNet.

About

• Few-shot setting: uses 1/3 of the defect images per category as references
• One-shot setting: uses only a single anomaly sample per defect category
• Adaptive Anomaly Mask and Texture Preservation modules ensure background consistency
• Can generate 1,000 anomaly image-mask pairs per category
• No additional training is required, so it is highly practical for real-world use.

See also

• Stable Diffusion V2.1
• ControlNet