AdaCLIP
Zero-shot anomaly detection that adapts CLIP with lightweight hybrid prompts across industrial and medical domains.
About
AdaCLIP, presented at ECCV 2024, takes the AnomalyCLIP idea a step further by adapting both the visual and text sides of CLIP with hybrid learnable prompts instead of text prompts alone. The extra parameters are tiny — about 4.6% of the frozen backbone — but the payoff is strong generalization across domains: the same model handles industrial defect datasets and medical anomaly datasets like skin lesions and colon polyps.
The authors report 94.6% image-level AUROC averaged across medical sets (with 89.6 max-F1) and 90.2% across industrial sets, and later comparison papers place it right alongside AnomalyCLIP on MVTec and VisA. The official implementation is MIT-licensed. As with all zero-shot anomaly detectors, it trades peak accuracy for breadth: supervised specialists still win on any single dataset they've been trained on.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (arXiv 2407.15795): medical avg image-level (94.6 AUROC, 89.6 max-F1); industrial avg (90.2 AUROC, 89.6 max-F1); DAGM 99.1 AUROC; Br35H (97.7, 92.4)
- Later comparison tables (AF-CLIP, AdaptCLIP) report consistent zero-shot numbers on MVTec/VisA
- No independent third-party replication found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER metrics, metrics-v1). Evidence: peer-reviewed ECCV 2024 paper (arXiv 2407.15795) with industrial + medical evals; MIT-licensed official repo (caoyunkang/AdaCLIP); consistent numbers in later comparison tables. Strengths: open, cross-domain results, lightweight adaptation. Docked for: (1) headline numbers are author-reported; (2) zero-shot accuracy still trails supervised specialists. License: MIT.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · ECCV 2024 paper arXiv 2407.15795; medical avg 94.6 AUROC / 89.6 max-F1 image-level; industrial avg 90.2 AUROC; DAGM 99.1 AUROC
- · MIT-licensed official repo caoyunkang/AdaCLIP; consistent numbers in later AF-CLIP/AdaptCLIP comparison tables
- · Only 4.6% added params over frozen ViT-L/14; all headline numbers author-reported, no independent replication