Skip to content
← Models
Open weightsIndependent researchers (ICLR 2024)/10 trustMetrics-checked

AnomalyCLIP

Zero-shot industrial anomaly detection and localization from CLIP with two learnable, object-agnostic prompts.

Visit official page →

About

AnomalyCLIP is one of the models that kicked off the CLIP-based zero-shot anomaly detection wave. Published at ICLR 2024, it learns just two text prompts — one for 'normal', one for 'anomalous' — that are deliberately object-agnostic, so the same model can flag defects on industrial parts it has never seen. Unlike earlier approaches that needed hand-crafted prompts or a separate decoder, it does both classification and pixel-level segmentation in one prompt-based pass.

The authors report 91.5% image-level and 91.1% pixel-level AUROC on MVTec AD, plus strong generalization to medical sets like BrainMRI and HeadCT. Later papers' comparison tables (AF-CLIP, GenCLIP) reproduce numbers in the same range, which is reassuring. The MIT-licensed official implementation makes it genuinely usable — the usual caveat being that zero-shot methods still trail supervised specialists like PatchCore on raw accuracy.

Specs

Context window

Not applicable — vision model; CLIP ViT-L/14 backbone at 518px input

Pricing

Free / open weights

Free quota

Fully open — no API

API access

No official API — self-hosted from the official MIT-licensed repo (zqhang/AnomalyCLIP)

Weights

Open weights

Released

Not disclosed

Benchmarks

  • Paper-reported (arXiv 2310.18961): MVTec AD image-level AUROC 91.5 / AP 96.2, pixel-level AUROC 91.1 / PRO 81.4; VisA image-level 82.1
  • Paper-reported medical: BrainMRI (90.3, 92.2), HeadCT (93.4, 91.6) image-level (AUROC, AP)
  • Corroborated by independent comparison tables in AF-CLIP and GenCLIP papers; supervised specialists (PatchCore, CDO) still score higher on MVTec

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

anomaly-detectiondefect-detectionindustrial-inspectionquality-controlmedical-imaging

Trust score — 7/10

/10 trustMetrics-checked

Scored 7/10 (TIER metrics, metrics-v1). Evidence: peer-reviewed ICLR 2024 paper (arXiv 2310.18961) with MVTec/VisA/medical evals; MIT-licensed official repo (675 stars); numbers corroborated by later papers' comparison tables. Strengths: open, independently re-evaluated, honest zero-shot positioning. Docked for: (1) headline numbers are author-reported; (2) supervised methods still beat it on MVTec, so 'zero-shot' is the value proposition, not raw SOTA. License: MIT.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
7.0/10
Maintenance · recency, releases, issue load
7.0/10
Openness · license + docs
9.0/10
Manual review · sanity read + red-flag check
7.0/10
  • · ICLR 2024 paper arXiv 2310.18961; MVTec AD 91.5 image / 91.1 pixel AUROC; VisA 82.1 image AUROC; medical HeadCT 93.4, BrainMRI 90.3
  • · MIT-licensed official repo zqhang/AnomalyCLIP (675 stars); numbers corroborated in AF-CLIP/GenCLIP comparison tables
  • · Supervised specialists (PatchCore, CDO) still score higher on MVTec — zero-shot is the value prop

Sources