Reproducing INP-Former: Reconstructing Intrinsic Normal Prototypes for Industrial Anomaly Detection

Yunfei Liang; Juan Pablo Osorio Rivera

cs.CVOct 31, 2026

Reproducing INP-Former: Reconstructing Intrinsic Normal Prototypes for Industrial Anomaly Detection

Authors: Yunfei Liang, Juan Pablo Osorio Rivera

Organizations: Faculty of Computer and Information Science University of Ljubljana

Abstract

We reproduce the central claims of INP-Former, a reconstruction-based method for multi-class industrial anomaly detection. INP-Former extracts Intrinsic Normal Prototypes (INPs) from each test image and uses them to guide a decoder that reconstructs normal features, without a stored memory bank. Using the authors' released code, we reproduce the reported performance on MVTec AD and VisA and extend the evaluation to MVTec AD 2, where image AUROC with the default settings is 63.4, compared with 99.7 on MVTec AD. On MVTec AD, a five-seed ablation shows that most of the performance comes from the INP mechanism, while the two auxiliary losses add small changes whose direction depends on the metric. A paired comparison with both losses disabled confirms the benefit of INP on MVTec AD 2. The default of six prototypes transfers less well: the plateau beyond four prototypes reproduces on MVTec AD and VisA, but on MVTec AD 2 a single prototype outperforms six in AU-PRO on the public split and on both hidden test splits. We also measure practical trade-offs in resolution, backbone size, training length, and LoRA adaptation, and examine performance by defect size, lighting condition, and a stricter AU-PRO limit. Several conclusions from the original benchmarks, including the default prototype count, do not carry over to harder data. Code: https://github.com/garasnote/Repro-INPFormer

Figures & tables

Claim	Question	Evidence in this study
Claim 1	RQ1	Main reproduction on MVTec AD and VisA (Section 4.1 ).
Claim 2	RQ1, RQ2	Loss and module ablation on MVTec AD (Section 4.1 ); paired INP-only versus No-INP comparison on the MVTec AD 2 public split (Section 5.3 ).
Claim 3	RQ1	Five-configuration loss ablation on MVTec AD (Section 4.1 ).
Claim 4	RQ1, RQ2	Prototype-count sweeps on MVTec AD and VisA (Section 4.2 ) and on MVTec AD 2 (Section 5 ).

Table 1: A table showing which of the research questions address which claims.

Configuration	Description
INP aggregation	$M$ learned queries cross-attend to the fused encoder features of each image, before the decoder bottleneck.
No-INP control	Bypasses aggregation and supplies the full pre-bottleneck encoder-feature sequence as decoder keys/values. Cross-attention remains; this is not vanilla self-attention, and equivalence to the original authors’ control is not established.
Soft mining	The exponent $y$ controls gradient reweighting of the cosine reconstruction objective. At $y=0$ , reweighting is disabled; reconstruction remains.
Coherence term	Mean nearest-prototype cosine distance over fused encoder features. $\lambda=0$ removes its contribution to the training objective.

Table 2: Configurations varied in our experiments.

Dataset	Role in this study
MVTec AD	15 categories of industrial products and textures; 3,629 normal training images and 1,725 test images. Original-benchmark reproduction ( Bergmann et al., 2019 ) .
VisA	12 categories of complex structures; 8,659 training images and 2,162 test images. Original-benchmark reproduction ( Zou et al., 2022 ) .
MVTec AD 2	8 categories, including Fabric; models trained on the AD2 training split and evaluated on the public test split ( Heckler-Kram et al., 2026 ) . The $M=1$ and $M=6$ submissions also have official private scores. Not evaluated in the original paper.
Real-IAD	Not included among our reproduced training or evaluation benchmarks ( Wang et al., 2024 ) due to training cost limitations.

Table 3: Datasets and evaluation scope. The prototype-count sweep uses the public MVTec AD 2 test split; official private scores are reported for

M=1

and

M=6

Setting	Value
Seed	1 for both runs.
Loss controls	$y=0$ and $\lambda=0$ in both runs; the reconstruction objective remains active. The two training configurations differ only in whether prototype aggregation is used.
Prototypes	$M=6$ for INP-only; the No-INP control has no aggregation step.
Training hardware	NVIDIA A100-SXM4-40GB for both runs.

Table 4: Settings specific to the paired INP-only and No-INP comparison on MVTec AD 2 (Section 5.3 ). Unlisted settings follow Section 3.3 .

Dataset	Source	I-AUROC	P-AUROC	P-AUPRO
MVTec AD	Original	99.7	98.5	94.9
MVTec AD	Ours	99.67	98.45	94.95
VisA	Original	98.9	98.9	94.4
VisA	Ours	98.90	98.97	95.27

Table 5: Multi-class anomaly detection on MVTec AD and VisA. “Original” indicates the results reported by Luo et al. (2025) ; “Ours” gives the reproduced results from our retraining of their models. This run is separate from the five-seed ablation in Table 6 . All entries are percentages.

Configuration	INP	$\mathcal{L}_{c}$	$\mathcal{L}_{sm}$	I-AUROC	P-AUPRO	I-AUROC	P-AUPRO
				Original		Ours
No INP	×	×	×	98.59	92.73	$98.19\pm 0.65$	$91.55\pm 1.19$
INP only	✓	×	×	99.53	94.69	$99.60\pm 0.03$	$95.14\pm 0.05$
INP + $\mathcal{L}_{c}$	✓	✓	×	99.61	95.10	$99.64\pm 0.02$	$95.23\pm 0.10$
INP + $\mathcal{L}_{sm}$ †	✓	×	✓	—	—	$99.64\pm 0.04$	$94.85\pm 0.08$
Full	✓	✓	✓	99.67	94.87	$99.62\pm 0.01$	$94.92\pm 0.07$

Table 6: Loss and module ablation on MVTec AD in percentage. “Original” values are from Table 5 of Luo et al. (2025) ; “Ours” are the means and standard deviations over five seeds. No-INP, INP+

\mathcal{L}_{sm}

, and Full come from an earlier set of runs than INP-only and INP+

\mathcal{L}_{c}

, which use seeds 1, 2, 3, 42, and 123.

	MVTec AD			VisA
$M$	I-AUROC	P-AUROC	AU-PRO	I-AUROC	P-AUROC	AU-PRO
1	98.37	96.54	92.21	97.34	97.66	93.00
2	99.36	98.12	94.39	98.15	98.53	93.80
4	99.55	98.39	94.75	98.81	98.84	95.04
6	99.67	98.45	94.95	98.90	98.97	95.27
8	99.60	98.42	94.90	98.84	98.98	95.23

Table 7: Prototype-count sweep on MVTec AD and VisA (percent). Results are from one seed. AU-PRO is integrated up to a false-positive rate of 0.30.

$M$	I-AUROC	P-AUROC	Public	Private	Priv.-mixed	[email protected]
			[email protected]
1	65.75	87.68	32.12	37.14	33.98	57.39
2	64.68	87.45	27.68	—	—	55.09
4	63.96	87.58	25.06	—	—	53.75
6	63.39	87.98	26.67	34.51	28.46	54.86
8	62.48	87.21	22.94	—	—	52.47

Table 8: Prototype-count sweep on MVTec AD 2 (percent), full model, seed 1. Public-split scores are computed locally. Private and private-mixed scores come from the official benchmark server, to which we submitted

M=1

and

M=6

only due to submission limits.

Configuration	I-AUROC	P-AUROC	[email protected]	[email protected]
No-INP control	59.90	86.08	18.98	48.20
INP-only	64.79	88.04	30.03	56.48
$\Delta$ (INP minus control)	+4.89	+1.96	+11.05	+8.29

Table 9: Paired INP-only and No-INP results on the eight-category MVTec AD 2 public test split (percent), seed 1. Both runs use 200 training epochs with the auxiliary losses disabled; reconstruction remains active. Training uses 448/392 resize/crop. Whole test images are resized to 392, with prediction maps restored to native size. INP-only uses

M=6

; No-INP feeds the full feature sequence to the decoder. The control retains decoder cross-attention, and its equivalence to the original authors’ control is not established. Deltas are calculated before rounding.

Backbone	Input/Crop	I-AUROC	P-AUROC	P-AUPRO
ViT-Base/14	224/196	99.21	97.97	92.07
ViT-Base/14	336/294	99.63	98.28	94.04
ViT-Base/14	448/392	99.67	98.45	94.95
ViT-Small/14	448/392	99.20	98.18	94.39
ViT-Base/14	448/392	99.67	98.45	94.95

Table 10: Resolution and backbone ablation on MVTec AD (percent). The ViT-Base/14 448/392 row is shown in both comparisons; it is the same run.

Config	Setting	FPS	Latency (ms)
Resolution (ViT-B, $M$ =6)	224/196	85.3	11.7
	336/294	50.1	20.0
	448/392	25.0	40.0
Backbone (448/392, $M$ =6)	ViT-Small/14	72.6	13.8
Backbone (448/392, $M$ =6)	ViT-Base/14	25.0	40.0
INP count (ViT-B, 448/392)	$M$ =2–16	25.0–25.2	39.7–40.0

Table 11: Model-forward inference speed on an NVIDIA L4 GPU at batch size 1 (50 warmup and 200 timed iterations). Timing excludes image preprocessing and output-map postprocessing.

	Trainable (%)		MVTec AD			VisA
Method	Total	Adapters	I-AUROC	P-AUROC	P-AUPRO	I-AUROC	P-AUROC	P-AUPRO
Zero-shot	—	—	69.15	79.92	57.17	51.47	72.60	31.60
LoRA $r=2$	4.71	0.14	97.22	96.21	91.30	91.79	96.29	85.11
LoRA $r=4$	4.85	0.29	97.81	96.62	91.87	93.17	96.79	87.55
LoRA $r=8$	5.12	0.58	98.16	96.96	92.47	94.44	97.25	89.91
Full train	—	—	99.67	98.45	94.95	98.90	98.97	95.27

Table 12: Parameter-efficient adaptation on MVTec AD and VisA (percent). “Zero-shot” is the Real-IAD checkpoint evaluated on the target dataset without adaptation; “Full train” is the 200-epoch model from Table 5 . “Total” counts the adapters, prototype tokens, and aggregation module. The encoder is frozen in all rows.

Epoch	Loss	I-AUROC	P-AUROC	P-AUPRO	$\Delta$ P-AUPRO
25	0.163	99.29	98.07	94.57	$-0.39$
50	0.145	99.50	98.26	94.76	$-0.20$
100	0.133	99.55	98.40	94.90	$-0.06$
150	0.128	99.59	98.45	94.95	$-0.01$
200	0.126	99.61	98.48	94.96	—

Table 13: Training convergence on MVTec AD. All metrics are means across 15 categories.

\Delta

shows the difference in percentage points from epoch 200 (final), separate from the results in Table 5 .

		I-AUROC change		[email protected] change
Condition	Categories evaluated	$M=1$	$M=6$	$M=1$	$M=6$
overexposed	8	+0.73	+1.11	-0.28	+0.49
underexposed	8	-2.61	-4.04	-3.05	-4.70
shift_1	8	-5.07	-1.72	-4.28	-9.26
shift_2	7	-0.24	-0.15	+0.46	-3.40
shift_3	7	-6.72	-2.53	-1.43	-5.90

Table 14: Change from regular lighting on the AD2 public split (percentage points), full models, seed 1. Negative values mean lower scores under the listed condition. Scores are computed within each condition.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Category	I-AUROC	P-AUROC	P-AUPRO
carpet	99.88	99.35	97.89
grid	99.92	99.42	97.75
leather	100.0	99.42	98.13
tile	100.0	97.68	88.36
wood	99.91	97.54	93.84
bottle	100.0	99.06	96.90

Appendix

Table 15: Per-category results on MVTec AD (ViT-Base/14, 448/392,

M

=6).

Category	I-AUROC	P-AUROC	[email protected]	[email protected]
can	+2.35	+1.32	-6.02	-0.22
fabric	+9.60	-2.85	+10.83	+17.92
fruit_jelly	-7.83	+4.09	+13.54	+11.29
rice	+3.62	-0.56	+13.92	-1.77
sheet_metal	+16.34	+3.90	+6.23	+6.56
vial	+3.37	-0.11	+27.42	+14.48

Appendix

Table 16: Per-category difference (INP-only minus No-INP) on the MVTec AD 2 public split, in percentage points. Both runs are shown in Table 9 ; negative values favor No-INP.

Category	I-AUROC	P-AUROC	[email protected]	[email protected]
can	50.65 / 52.47	64.97 / 69.36	2.33 / 1.17	22.14 / 24.82
fabric	56.65 / 69.16	83.47 / 90.74	21.91 / 22.87	43.88 / 51.98
fruit_jelly	74.58 / 61.00	94.87 / 95.57	39.40 / 30.44	75.45 / 74.84
rice	70.42 / 68.76	93.93 / 93.51	45.85 / 37.11	67.05 / 60.25
sheet_metal	62.50 / 54.77	89.04 / 89.11	13.25 / 7.74	30.98 / 22.82
vial	91.65 / 77.39	90.14 / 85.76	65.40 / 54.72	89.39 / 83.30

Appendix

Table 17: Full-model AD2 public scores (percent), seed 1, with both auxiliary losses enabled. Each cell is

M=1

M=6

. These category scores underlie Table 8 ; they are distinct from the loss-disabled architecture comparison.

Method	Input size	Private	Private-mixed
EfficientAD-S	$256\times 256$	30.8	19.2
PatchCore	$256\times 256$	28.8	26.0
EfficientAD-S	$512\times 512$	40.0	25.3
PatchCore	$512\times 512$	41.9	35.8
INP-Former, $M=1$	$392\times 392$	37.14	33.98
INP-Former, $M=6$	$392\times 392$	34.51	28.46

Appendix

Table 18: Hidden-split [email protected] (percent). Baseline scores are reported by Heckler-Kram et al. (2026) for EfficientAD-S ( Batzner et al., 2024 ) and PatchCore ( Roth et al., 2022 ) . INP-Former scores are our official server results. Methods use different training settings and input sizes.

Condition	$M$	I-AUROC	P-AUROC	[email protected]	[email protected]
overexposed	1	69.66 / 70.39	88.39 / 88.08	34.43 / 34.15	58.24 / 58.21
	6	68.21 / 69.33	89.98 / 89.60	32.75 / 33.25	58.16 / 58.01
underexposed	1	69.66 / 67.05	88.39 / 87.14	34.43 / 31.38	58.24 / 56.40
	6	68.21 / 64.18	89.98 / 87.42	32.75 / 28.06	58.16 / 54.46
shift_1	1	69.66 / 64.59	88.39 / 88.19	34.43 / 30.15	58.24 / 57.38
	6	68.21 / 66.49	89.98 / 89.07	32.75 / 23.50	58.16 / 55.89

Appendix

Table 19: Matched regular / changed-lighting means (percent), with

M=1

and

M=6

adjacent. Each metric cell is regular / changed. Conditions use the same matched categories as Table 14 ; its deltas were computed before rounding.

Category	Condition	$n$	I-AUROC	P-AUROC	[email protected]	[email protected]
can	regular	12+15	51.11 / 56.11	65.52 / 72.00	0.80 / 1.45	20.68 / 28.21
	overexposed	12+15	47.22 / 54.44	66.34 / 71.65	1.54 / 1.69	21.98 / 28.48
	shift_1	12+15	56.67 / 69.44	66.90 / 70.67	5.60 / 2.11	27.29 / 28.54
	shift_2	12+15	51.11 / 45.56	62.56 / 66.65	0.80 / 0.31	18.41 / 18.57
	shift_3	12+15	55.56 / 53.89	66.48 / 69.36	4.33 / 0.77	26.54 / 23.90
	underexposed	12+15	42.22 / 66.11	62.32 / 68.78	0.13 / 0.86	17.76 / 22.79

Appendix

Table 20: Condition-level AD2 public scores (percent). Each metric cell is

M=1

M=6

;

n

is the normal + anomalous image count shared by both models. Normal images are condition-matched. Conditions absent from a category are unavailable, not zero.

Anomaly Detection

Anomaly detection finds rare or unexpected patterns in data, from defective products and system faults to fraud and security threats.

See all 681 papers

Flownomaly: Single-Step Conditional-Mean Recovery for Few-Shot Anomaly Detection
Oct 23, 2026Bernardo Perrone Ribeiro, Erkut Akdag, Egor BondarevFew-Shot Anomaly Detection Industrial Anomaly Detection
AIMS Group, Department of Electrical Engineering, Eindhoven University of Technology
BoRAD: Bootstrap your Own Representations for Multi-class Anomaly Detection
Jun 12, 2026Duy Hoang Khuong, Tri Nguyen Minh, Ngu Huynh Cong VietRepresentation Learning Industrial Anomaly Detection
Department of Artificial Intelligence FPT University Ho Chi Minh City, Vietnam · Department of IT FPT University Ho Chi Minh City, Vietnam · Department of Computing Fundamental FPT University Ho Chi Minh City, Vietnam
What Remains Normal? Clean Images Miss Useful Near-Defect Normal Patches for Anomaly Detection
Aug 24, 2026Joongwon Chae, Runming Wang, Peiwu QinIndustrial Anomaly Detection Reconstruction-Based Anomaly Detection
Institute of Biopharmaceutical and Health Engineering, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · RatelSoft · Guangdong Provincial Laboratory of Traditional Chinese Medicine, Hengqin, Guangdong, China

Oct 23, 2026J/K move · Enter open · S save

Flownomaly: Single-Step Conditional-Mean Recovery for Few-Shot Anomaly Detection

Bernardo Perrone Ribeiro, Erkut Akdag, Egor Bondarev

AIMS Group, Department of Electrical Engineering, Eindhoven University of Technology

Few-Shot Anomaly Detection Industrial Anomaly Detection Anomaly Localization Reconstruction-Based Anomaly Detection

Open paper AskPDF arXiv

Few-shot industrial anomaly detection requires modeling category-specific normality from only a handful of defect-free images. Common approaches tie inference to stored exemplars, impose rigid linear structure, or risk reproducing anomalous content. Flownomaly instead casts the problem as normal-counterpart recovery in frozen DINOv2 feature space. Given normal supports, a velocity field is learned through conditional flow matching over frozen patch tokens to transport Gaussian noise toward normal features. Along straight paths, a single recovery step using one evaluation of the optimal field yields the posterior mean of the clean normal token without trajectory integration. The operating time governs how strongly recovery modifies a query token. With no anomalous validation samples available for tuning, a normal-only Jacobian gain-to-noise criterion selects this time from the supports alone. Squared residuals between query tokens and their recovered counterparts yield anomaly signals for both dense localization and image-level detection. Across 1-, 2-, and 4-shot benchmarks on MVTec-AD and VisA, together with sample-shot evaluations on Real-IAD, Flownomaly achieves state-of-the-art performance against recent category-specific methods. In the 1-shot setting, image- and pixel-level AUROC reach 97.8% and 98.0% on MVTec-AD and 95.3% and 98.7% on VisA, with a 2.1-point image-level AUROC improvement over the strongest baseline on VisA. Code is available at https://github.com/b-rbmp/Flownomaly.

Abstract

Anomaly Detection

Explore similar work

Flownomaly: Single-Step Conditional-Mean Recovery for Few-Shot Anomaly Detection

BoRAD: Bootstrap your Own Representations for Multi-class Anomaly Detection

What Remains Normal? Clean Images Miss Useful Near-Defect Normal Patches for Anomaly Detection