AnomalyCLIP
AnomalyCLIP is a zero-shot anomaly detection method that learns object-agnostic prompts to detect and localize anomalies in unseen categories.
View on GitHub ↗Detection setup
If you upload multiple images, one is selected at random.
Uses 6 normal references: repeats uploaded images if fewer than 6, or randomly selects 6 if more.
Uses 6 abnormal references with the same repeat or random selection rule.
If omitted, the app looks for <stem>_mask next to each abnormal image; any missing mask falls back to a full-image mask. Default weights: IDEAL_n6a6.pth.
Upload from any machine connected to this server. Each object subfolder contains test images. For InCTRL / IDEAL, include a fewshot/ folder with per-object references.
AnomalyCLIP is a zero-shot anomaly detection method that learns object-agnostic prompts to detect and localize anomalies in unseen categories.
View on GitHub ↗FAPrompt is a zero-shot anomaly detection method that learns fine-grained abnormality prompts and adapts them to each image to detect subtle defects.
View on GitHub ↗InCTRL is a few-shot anomaly detection method that uses normal reference images as in-context prompts to identify anomalies in a test image.
View on GitHub ↗IDEAL is a few-shot anomaly detection method that learns intrinsic deviation patterns from normal and abnormal references to detect anomalies.
View on GitHub ↗