AFLAD — Reproduction Package

Annotation-Free signal fusion for Logical and structural Anomaly Detection on MVTec LOCO AD.

This repository exists so that the numbers in the paper can be recomputed, not taken on trust. It ships the per-image anomaly scores that the reported results are built from.

Main setting — image-level AUROC (L+S) = 0.9723 ± 0.0011 (seeds 42/43/44) logical 0.9685 · structural 0.9762 · +2.54 pp over the same-harness SALAD reproduction (0.9469)

---

1. Verify in 30 seconds

The headline is recomputed from per-image scores alone. No 18 GB feature cache, no 9 GB of checkpoints, no GPU — numpy and scikit-learn suffice.

python3 code/analysis/reproduce_from_package.py --pkg .

Windows: the JSON artifacts contain Korean text, so run under UTF-8 — set PYTHONUTF8=1 (or python -X utf8 ...). The scripts pass encoding="utf-8" explicitly, so this is a safety net rather than a requirement. File hashes in MANIFEST.md are computed on LF line endings; .gitattributes pins that, so a Windows clone with core.autocrlf=true still matches.

  seed 42: L+S 0.9721 (logical 0.9662 / structural 0.9780)
  seed 43: L+S 0.9714 (logical 0.9682 / structural 0.9745)
  seed 44: L+S 0.9735 (logical 0.9710 / structural 0.9759)

  3-seed L+S = 0.9723 ± 0.0011
    LS         reproduced 0.9723  paper 0.9723  diff 0.00001  PASS
    logical    reproduced 0.9685  paper 0.9685  diff 0.00005  PASS
    structural reproduced 0.9762  paper 0.9762  diff 0.00004  PASS

That script is the entire test-free protocol: z-normalise each branch score using the mean and standard deviation of the normal samples of the official validation split, sum with equal weights (1,1,1), compute AUROC per defect type, and report L+S = (logical + structural) / 2. No log scaling, no robust statistics, no clipping.

Why ship scores rather than code alone

Re-training with the same code, the same pseudo-labels and the same seed does not reproduce the branch scores. We measured this with a replicate run (same machine, seed 42, independent re-training):

what differs	Spearman
seed only	0.9381
machine only	0.9378
run only (nothing else)	0.9388

Changing nothing at all produces the same divergence as changing the seed and the machine. The cause is cudnn.benchmark = True in train_normal_unet.py: convolution algorithms are selected by wall-clock timing, and many of the selected kernels use atomicAdd in the backward pass. The seed fixes initialisation — weights of a module that never runs are bit-identical across runs — but it does not fix the arithmetic.

The fused number does reproduce. A branch-level gap of −0.0026 shrinks to −0.0007 after fusion, which is 0.61× the seed standard deviation (0.0011). So the precise reproducibility claim of this method is "the fused number reproduces", not "the branch scores reproduce" — and checking that requires the score files themselves.

Evidence: results/tables/psad_variance_decomp.json, results/tables/psad_3way_compare.json.

---

Environment

The verification in §1 needs only Python 3 with numpy and scikit-learn (requirements-verify.txt). The shipped scores were produced in a GPU container with PyTorch 2.11.0 (CUDA 12.8), timm 1.0.22 and OpenCV 4.13.0; the branch scripts under code/ assume that environment plus the upstream implementations pinned in §8.

---

2. Method

A sum of three branch scores. Each catches a different failure mode.

Branch	What it measures	Code
Reconstruction — EfficientAD-S	local texture / structure deviation	code/branches/reconstruction_ead/
Patch memory — DINOv3-L L17 PatchCore	distance from the normal patch manifold	code/branches/patch_memory_pc/
Composition — PSAD-hc	component area histogram (h) + appearance embedding (c)	code/branches/composition_psad/

Component labels for the composition branch come from a U-Net segmenter distilled from CSAD pseudo-labels, with two improvements, both adopted on train/val evidence only:

• Automatic merging of confusable class pairs — pairs with C[i,j] + C[j,i] > 0.10 in the confusion matrix between train pseudo-labels and U-Net predictions are merged. Across all five categories only screw_bag c4↔c5 qualifies (0.168 / 0.261 / 0.142 over three seeds; the runner-up is ≤ 0.012 everywhere — a 10× margin, so a single threshold isolates that pair without per-category hand-tuning).
• Rotation TTA — inference at four 90-degree rotations, which lie within the rotation range of the training augmentation, averaged. The rule "must reduce both the mean and the maximum of the validation-normal score tail" was applied identically to all five categories; only screw_bag passed.

---

3. Paper element → artifact → code

Paper element	Artifact	Producing code
Main headline 0.9723	results/main/testfree_final_eadfix_3seed.json	code/fusion/prereg_testfree.py --protocol paper --kind hc --psad-tag _tta_merge
Legacy-protocol variant 0.9733	results/main/testfree_final_eadfix_origproto_3seed.json	same script (seed-averaged EAD·PC)
hcp variant (appendix)	results/main/testfree_final_hcp_3seed.json	same script, --kind hcp --psad-tag _tta_merge_k5
Baseline comparison table	results/tables/table5_3seed_42_43_44.json	code/baselines/realign_*.py
PatchCore family (train-only bank)	results/tables/patchcore_family_trainonly_3seed.json	code/branches/patch_memory_pc/patchcore_family_trainonly.py
EAD-M with full ImageNet	results/tables/realign_eadm_fullimagenet_3seed.json	code/baselines/realign_eadm_fullimagenet.py
Branch-alone / ablation	results/tables/testfree_supplements_3seed.json	code/analysis/testfree_supplements.py
Deployment configurations	results/tables/deployment_operating_point.json	code/analysis/deployment_operating_point.py
Requirement–cost curve	results/tables/requirement_cost_curve.json	code/analysis/requirement_cost_curve.py
Subtype × four score sources	results/tables/subtype_branch_auroc.json	code/analysis/subtype_branch_auroc.py
ComAD per-axis seed spread	results/tables/comad_axis_perseed.json	code/analysis/comad_axis_perseed.py
Inference parameter counts	results/tables/param_counts_inference.json	code/analysis/measure_param_inference.py
Forward-execution measurement	results/tables/forward_profile.json	code/analysis/profile_forward_modules.py
EAD-S 5-seed reproduction check	results/reproduction/ead_5seed_verdict.json	code/audit/ead_5seed_verdict.py
Pre-registration of design choices	results/diagnostics/PREREG_branch_improvement.md	—
Accept / reject evidence	results/diagnostics/criterion_*.json, sweep_*.json	code/selection_criteria/

Directory roles and seed coverage for the per-image scores: results/per_image_scores/README.md.

File names keep their internal numbering; the paper's tables map to them as follows.

Paper	Artifact
Table 2 (same-harness baselines)	results/tables/realign_3seed_42_43_44.json, results/tables/realign_eadm_fullimagenet_3seed.json
Table 3 (branch alone / fusion per category)	results/tables/testfree_supplements_3seed.json, results/figures/testfree_pairwise_percat_3seed.json
Table 4 (axis decomposition)	results/tables/branch_axis_decomposition.json
Table 5 (deployment arrangements)	results/tables/requirement_cost_curve.json, results/tables/deployment_operating_point.json
Tables C1–C2 (per seed)	results/main/testfree_final_eadfix_3seed.json, results/tables/table5_3seed_42_43_44.json
Table A1 (localization)	results/tables/p18_heldout_fullres.json, results/tables/p21_pixel_auroc.json
Tables D1–D2 and Figure 8 (double dissociation)	results/tables/tableC1_full_grid.json, results/tables/tableC2_true_cardinality.json
Appendix D subtype table	results/tables/subtype_branch_auroc.json

---

4. What "test-free" means here

No configuration was ever selected using the test split. Decisions used two criteria only.

Criterion	What it measures	Data
train discriminability	can the branch separate logical anomalies from normals	perturbations over train pseudo-labels + leave-one-out 1-NN AUROC
val stability	does z-normalisation survive outliers	max |z_robust| of the validation-normal score tail

The two can disagree. A single validation-normal image that inflates the standard deviation shrinks that branch's test z-scores across the board, effectively switching it off. Two candidates were rejected for exactly this reason (k-NN with k=5; a cardinality distance).

The highest-scoring candidate was also rejected. An h:c weighting (w_c = 0.25) gave the largest criterion improvement (+0.0150 mean), but because the criterion manipulates masks, the area term changes directly while the appearance term changes only indirectly — so the criterion structurally underrates the appearance term. A criterion cannot select on something it cannot measure. Of 11 candidates, 2 were adopted and 9 rejected; the record is in results/diagnostics/PREREG_branch_improvement.md.

Easy to misread: code/selection_criteria/soft_areas.py iterates over the test split. That is scoring, not selection — that experiment (soft masks) was rejected on the validation tail (pushpins 1.77 → 10.03). Test AUROC is used only for the single final report.

---

5. What is not included, and why

Item	Size	How to regenerate
DINOv3-L feature cache	~18 GB	code/branches/patch_memory_pc_extra/I_dinov3_sl_cache.py
U-Net segmenter checkpoints	~9 GB	code/branches/composition_psad/train_unet_seeds.sh
EfficientAD training weights	~2 GB	code/branches/reconstruction_ead/train_ead_seeds.sh
MVTec LOCO AD dataset	~6 GB	from MVTec Software GmbH
DINOv3-L/16 weights	—	Hugging Face facebook/dinov3-vitl16-pretrain-lvd1689m (gated license; accept the terms, then download)
Rejected exploratory PSAD variants	—	flags in score_psad.py

Per-image scores are included, because they are sufficient for the verification in §1. The bulk items above are needed only to rebuild those scores from scratch — and, per §1, rebuilding does not reproduce them exactly.

Not available: the human-annotation variant (5 categories × 3 seeds were never produced, so the corresponding column was removed from the paper).

---

6. Paths

Scripts are shipped exactly as they ran, so the repository root is hardcoded in most of them. Execution records were not edited after the fact. To retarget:

grep -rl /workspace/ai-vision-research . | xargs sed -i "s|/workspace/ai-vision-research|$PWD|g"

code/analysis/reproduce_from_package.py contains no absolute path and runs anywhere. Details in code/RETARGET.md.

---

7. Integrity

MANIFEST.md lists every file with its origin path and SHA-256 prefix. code/audit/ contains the scripts that built and checked this package, including a coverage audit that extracts every numeric value from the manuscript and confirms it is backed by an artifact here.

License

MIT for our code and derived artifacts — see LICENSE. Third-party components (dataset, model weights, upstream implementations) remain under their own licenses and are not redistributed; they are listed in THIRD_PARTY_NOTICES.md.

8. Upstream implementations

Two branches follow official implementations, and the composition branch uses CSAD's released pseudo-labels. Pinned commits:

Component	Repository	Commit
efficient_ad_official	https://github.com/nelson1425/EfficientAD.git	fcab514
PSAD_official	https://github.com/oopil/PSAD_logical_anomaly_detection.git	1ae3146
CSAD_official	https://github.com/Tokichan/CSAD.git	d46cf85

Only derived artifacts are redistributed here; the upstream code and any model weights must be obtained from those repositories.