KIMJINYOUNG/TED-reproducibility

TED reproducibility: research artifacts and verified inference

Research artifact release for TED: Text-Axis Evidence Decomposition for Prompted Anomaly Localization, arXiv:2609.39033.

Eleven pilab configurations are selectable: FAPrompt L/14-336, captured FAPrompt B/16+ seed1 and H/14 seed0, AA-CLIP main/source-limit-1, AdaptCLIP OpenAI/L/14-336, RawCLIP OpenAI L/14/H14/L336, and ImageBind. The ten-worker routing audit passes all42 fixed-image/recorded-strength comparisons. ImageBind separately passes three original-code Docker HTTP comparisons and three direct-versus-selector cases. The previous ten registry entries and gateway source are unchanged. Each package has its own documented validation scope. The published verified-results index covers23 full-dataset configurations; it does not establish every paper mean, standard deviation or ablation.

Saved checkpoints and banks: download now

All 11 historical host-weight objects and 114 source-bank objects are already public. Their original paths and recipe/seed/transfer bindings are preserved. Use SAVED-ARTIFACTS.md for four pinned downloads and ready-to-run acquisition commands; these saved-input downloads do not need to wait for ongoing fresh GPU evaluations.

Captured ImageBind release

The full BTAD replay matches all12 archived metrics at two decimals. The package retains source, backbone and captured source-bank/calibrator state. Fresh archive extraction verifies all930 inputs and three CPU image comparisons; the pilab Docker worker preserves all three maps and raw scores in three original-code HTTP comparisons. See IMAGEBIND-CAPTURED-SERVING.md.

Captured RawCLIP OpenAI release

The MVTec-to-BTAD configuration matches all12 archived Base/T-TED/C-TED metrics at two decimals. Its public package preserves the original source bank, source calibrators, source files and pinned backbone, with no dataset images or serving-time refitting. See RAWCLIP-CAPTURED-SERVING.md.

The pinned GitHub acquisition client downloads from immutable revision b1baf4e0ef3e15affcc0bb81996b41626c7b4fc7 and verifies the archive size/SHA and all930 extracted files:

python -m reproduction.rawclip_release openai-l14 NEW_RAWCLIP_BUNDLE --cache NEW_CACHE

An anonymous public acquisition into a new cache and extraction directory verified the complete archive and all930 files. Base, T-TED and C-TED maps and fused image scores then matched the original output block in three protocol-fixed CPU image comparisons with original workspace/model reads and unrelated network connections denied; and the pilab Docker worker passes the same three original-code HTTP comparisons. These image checks are separate from full-dataset metrics. Image scores are raw fused scores, not probabilities.

Captured FAPrompt releases: verified public acquisition

The H/14 seed0 and B/16+ seed1 BTAD executions each match eight archived per-seed metrics at two decimals. Their archives retain actual source banks, fitted branch calibrators, original prepared source, checkpoint and backbone. No calibrator is refitted for serving and no dataset images are included.

Both archives are public at immutable revision f49bcc6c857606fd7edb1148856157532f6b2271. Each was downloaded anonymously with standard-library HTTPS, extracted into a new directory, and verified against size/SHA plus all932 input files and terminal state bindings. CPU image inference matches the original evaluator output block on three canonical BTAD images at all three recorded strengths (0.5,1.0,1.5): nine cases per variant with zero map/raw-score error. Original workspace/model reads and network access are denied during these inference checks.

Use the GitHub package's pinned client with an empty cache and new destination:

python -m reproduction.faprompt_release bplus-seed1 NEW_BPLUS_BUNDLE --cache NEW_CACHE
python -m reproduction.faprompt_release h14-seed0 NEW_H14_BUNDLE --cache NEW_CACHE

See FAPROMPT-CAPTURED-SERVING.md and the immutable GitHub checks. This validates these specific packages and image cases; all-dataset/seed means, other models and standalone dependency images remain incomplete.

Additional historical model checkpoints (2026-10-08)

The eleven exact host checkpoints for 207 traced AA-CLIP, FAPrompt, AdaptCLIP,

AdaCLIP, and BayesPFL recipes are now preserved in host-checkpoints-20261008.tar.

See HOST-CHECKPOINTS.md for extraction, byte verification,

provenance, and component licenses. host-checkpoints-archive.json pins the

archive, and host-checkpoints.json maps each recipe to its objects.

source.zip and source-manifest.json preserve 920 research source files.

This addition is a historical artifact release, **not completion of all-dataset

reproduction or all-model API deployment**. Banks, backbone assets, calibrators,

full clean runs, and portable execution remain in progress. The FAPrompt inference

release below is retained with its original validation scope.

Additional adapted-host source banks

host-source-banks-20261008.tar preserves 92 hash-verified bank objects mapped to the 207 adapted-host recipes. See SOURCE-BANKS.md for byte verification and the distinction between explicit historical paths and replayed cache-selection evidence. This preserves source features; it does not claim fresh bank rebuilding or complete all-dataset reproduction. Raw-backbone and ablation banks are released in the separate archives described below; fresh rebuilding and complete reruns remain pending.

Legacy FAPrompt L/14-336 files and provenance

• host.pth: preserved FAPrompt checkpoint trained_on_mvtecad/epoch_15.pth; not a newly trained checkpoint.

• source_bank_mvtec_seed0.pt: preserved source bank with 4,096 false-positive and 4,096 defect features, scores and interface metadata. These are features, not dataset images.

• calibrator.pt: two branch calibrators captured directly during the historical GPU recipe, projected source banks and inference settings. No target labels enter calibration.

• training_recipe.json: historical arguments and deployment settings.

• capture_historical_calibrators.py: the capture script as executed in the research workspace; contains workspace paths and is an audit record, not a portable full training entry point.

• runtime-source.zip: FastAPI inference engine, vendored evaluator, tests, Dockerfile and Compose recipe.

• backbone.json: upstream OpenAI CLIP download URL and exact digest. The 934 MB backbone is fetched separately.

• MANIFEST.json: SHA-256 and byte lengths for release files.

The calibration source is MVTec AD, and the evaluation target is BTAD. Source anomaly labels are used; this is not training from normal images only. Neither host pretraining nor fresh mining of the archived source bank has been reproduced in this release.

Validation scope

The preserved FAPrompt paper recipe was rerun on all 741 BTAD test images with historical batch size 4. Base and C-TED I-AUROC, P-AUROC, PRO and pixel AP matched their archived results at two decimal places. C-TED alpha 0.5 values (%): 89.24 / 94.99 / 70.95 / 45.01.

The broader provenance audit linked 300/300 main-table means to archived summaries, and 627/640 host-table means. The published verified-results index identifies the configurations independently checked in full-dataset reruns. This is not a claim that every paper experiment has been rerun.

The HTTP API processes one image at a time. Historical batch behavior and host postprocessing are retained in the evaluator; paper batch-4 benchmark scores must not be presented as measurements of the batch-1 CPU API. See included release validation reports for actual API checks.

Run the legacy FAPrompt L/14-336 release with Docker

Download this repository, then use Python 3.10+:

python prepare_models.py

TED_MODEL_DIR="$PWD" docker compose -f runtime/compose.yaml up -d --build

curl http://127.0.0.1:18080/ready

curl -X POST http://127.0.0.1:18080/predict \

  -H 'Content-Type: image/png' --data-binary @your-image.png

Open http://127.0.0.1:18080/ for the upload demo or /docs for the API. CPU inference is intended for a small research demo; no throughput SLA is claimed. The API returns raw host/C-TED maps and an unchanged host image score, which is not a defect probability. No operational pass/fail threshold has been calibrated.

The generic standalone CPU image was previously tested with offline model loading. The pilab deployment uses the host-runtime Compose variant because its shared root Docker disk has insufficient space for the standalone image. The production machine's Python environment is mounted read-only; this deployment is not yet fully portable. GPU Compose is a recipe only and must obey the host's scheduler policy.

Attribution and usage

FAPrompt: Jiawen Zhu, Yew-Soon Ong, Chunhua Shen and Guansong Pang, Fine-grained Abnormality Prompt Learning for Zero-shot Anomaly Detection, ICCV 2025. Upstream code, MIT notice in LICENSE-FAPrompt.txt and inside the runtime archive.

MVTec AD: Paul Bergmann, Michael Fauser, David Sattlegger and Carsten Steger, MVTec AD — A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection, CVPR 2019. Dataset and license. Source-derived artifacts retain the dataset's CC BY-NC-SA 4.0 restrictions; see LICENSE-SOURCE-DATA.txt. Use for non-commercial research and portfolio demonstration. This repository does not grant a blanket commercial license to upstream artifacts.

OpenAI CLIP: upstream repository. Third-party notices remain applicable. Original datasets, credentials, and private conversation logs are not included. The AA serving bundles include their exact pinned CLIP backbone weights; upstream notices remain applicable.

Frozen-backbone bank archive

raw-source-banks-20261008.tar preserves 10 additional source banks for the 25 traced RawCLIP/ImageBind main-table configurations. See RAW-SOURCE-BANKS.md for verification and limitations. Full GPU reproduction, MVTec AD 2 access, fresh bank rebuilding, and all-model deployment remain unfinished.

Weak-source ablation bank archive

weak-source-banks-20261009.tar preserves the 12 source banks for all 28 archived weak-source configurations. See WEAK-SOURCE-BANKS.md for evidence levels and verification. Fresh GPU ablation replay and full-paper deployment remain unfinished.

AA-CLIP serving bundles (October 9)

See AA-SERVING.md and AA-SERVING-ARCHIVES.json for two complete serving-input archives, including original weights, prepared research source, captured calibration/bank state, run settings and execution evidence. Neither contains dataset images. Both recipes matched 8/8 archived full-BTAD seed-0 metrics on A10; main and weak-source are distinct configurations. Three-image same-pilab original-equation/Docker parity passed for both. These HTTP comparisons use the same CPU environment for original equations and serving.

The earlier three-worker gateway check passed all nine direct-worker/gateway image predictions. The current selector has eleven configurations, with the later ten-worker42-case audit and three new ImageBind cases documented separately above.

Both AA serving archives have now passed anonymous download, new-directory full-input verification and isolated CPU three-image raw-map parity. See AA-SERVING.md for pinned acquisition commands and precise evidence scope.

RawCLIP H/14 and L/14-336 now also pass fresh anonymous archive download, 930 newly extracted inputs each and three original-code CPU image comparisons each. Public acquisition and replay evidence. FAPrompt VisA bank32 seed0 has also passed the full2162-image reported protocol and all8 archived metric comparisons; verified index.