Modelpublic

cubert-gmbh/adaclip

sourceHugging Faceunknownupdated 18d agoView on Hugging Face

0likes

Model Card

AdaCLIP pretrained prompt weights (mirror)

Byte-identical mirror of `caoyunkang/AdaCLIP` (weights published through Google Drive links in the project README) for the Cuvis.AI plugins. No fine-tuning, conversion or re-serialization: the files below are the upstream release as published, with the upstream licence text in LICENSE.

Why this mirror exists

Cuvis.AI provisions model weights once into a shared cache (download-model download <name> from cuvis-ai-core) and then runs its pipelines in an offline, token-free runtime that only reads that cache. Hosting the exact upstream files under the cubert-gmbh organisation makes that provisioning reproducible (commit-pinned and sha256-verified) and takes the per-user Hugging Face account and token out of the users' path.

Files and provenance

File	Upstream source	Upstream revision / id	sha256	Size
`pretrained_all.pth`	Google Drive file `1Cgkfx3GAaSYnXPLolx-P7pFqYV0IVzZF`	`1Cgkfx3GAaSYnXPLolx-P7pFqYV0IVzZF`	`33e8d3db1cb4aab030866b8b70a46e10aa27ebf2c23b5463cb07f2574addd98c`	42.7 MB
`pretrained_mvtec_colondb.pth`	Google Drive file `1xVXANHGuJBRx59rqPRir7iqbkYzq45W0`	`1xVXANHGuJBRx59rqPRir7iqbkYzq45W0`	`be51a42c052bd4cf060e54f503a1f5d0b2a3b899bc8dc2e243042f18b215427e`	42.7 MB
`pretrained_visa_clinicdb.pth`	Google Drive file `1QGmPB0ByPZQ7FucvGODMSz7r5Ke5wx9W`	`1QGmPB0ByPZQ7FucvGODMSz7r5Ke5wx9W`	`3deabbbaf1e412cfdfcb42923a500b986f4b9ee96ccbc7a735d89dbc87df44c8`	42.7 MB
`LICENSE`	<https://raw.githubusercontent.com/caoyunkang/AdaCLIP/main/LICENSE>	n/a (URL; sha256 pinned)	`58bf3cbb252fb8ee158f71b5eefa0f93e24632f587926659eb2638aa0df6c618`	1.1 kB

Mirrored on 2026-09-04 by Cubert GmbH from the sources above. The sha256 values are the upstream values; tools/mirror_weights.py check in cuvis-ai-core re-verifies this repository against them and against the upstream licence text.

Licence

The AdaCLIP code and these released checkpoints come from the AdaCLIP project (Yunkang Cao et al., "AdaCLIP: Adapting CLIP with Hybrid Learnable Prompts for Zero-Shot Anomaly Detection", ECCV 2024, <https://github.com/caoyunkang/AdaCLIP>). The project publishes its code under the MIT License (LICENSE, verbatim). The authors published these weights through Google Drive links in the project README without a separate licence statement; they are redistributed here unchanged, as released, so that Cuvis.AI can provision them without Google Drive.

The license: unknown tag reflects that no licence statement covers the weights; the LICENSE file is the project's code licence. The checkpoints were trained on auxiliary anomaly-detection datasets (MVTec AD, VisA, ClinicDB, ColonDB). MVTec AD is licensed CC BY-NC-SA 4.0; check the training-data licences for your use. Naming: the upstream README's weights table labels the Drive file pretrained_mvtec_colondb.pth as "MVTec AD & ClinicDB" and pretrained_visa_clinicdb.pth as "VisA & ColonDB", while its Train section pairs MVTec AD with ColonDB and VisA with ClinicDB, matching the file names. This mirror keeps the upstream file names and renames nothing. The AdaCLIP authors have been notified of this mirror. If you are a rights holder and object to this redistribution, open a discussion on this repository and the files will be taken down.

Usage with Cuvis.AI

text

uv run download-model download adaclip_all
uv run download-model download adaclip_mvtec_colondb
uv run download-model download adaclip_visa_clinicdb

download-model (from cuvis-ai-core) provisions the file into the shared Hugging Face cache and verifies its sha256. From cuvis-ai-core 0.16.0 on, ModelWeights.resolve("<name>") returns the cached path and the cuvis-ai-adaclip nodes resolve their weights through it, so a pipeline needs no manual step once the weights are provisioned; earlier plugin releases still fetch from their original upstream sources.