RealIAD-HD: High-Res Industrial Anomaly Benchmark

• The paper introduces RealIAD-HD as a high-resolution benchmark derived from Real-IAD, instantiated as RealIAD-2K with standardized 2048x2048 top-view images from three product categories.
• It focuses on detecting subtle anomalies with very small defect areas (<0.01%), challenging methods to preserve fine-grained discriminative details.
• The benchmark employs multi-resolution feature fusion and a dual-branch architecture to mitigate information loss from conventional downsampling.

RealIAD-HD denotes a high-resolution industrial anomaly detection benchmark derived from Real-IAD, the large-scale real-world multi-view industrial anomaly detection dataset. In the high-resolution anomaly detection literature, the concrete experimental instantiation is RealIAD-2K: a standardized top-view subset comprising Bottle Cap, Mint, and USB Adaptor, with all images resized to $2048\times 2048$ (Zhang et al., 18 Aug 2025). The term is not introduced as a named variant in the original Real-IAD paper, where high resolution is an inherent property of Real-IAD itself rather than a separately defined release (Wang et al., 2024).

1. Definition and nomenclature

The term has two distinct usages in the published material considered here.

Usage	Source	Meaning
RealIAD-HD	HiAD benchmark paper	Curated high-resolution benchmark derived from Real-IAD; experimentally instantiated as RealIAD-2K
“RealIAD-HD”	Real-IAD-MVN explanatory material	Informal label for the high-precision, multi-view pseudo-3D variant Real-IAD-MVN

In the original Real-IAD paper, there is no dataset variant formally named “RealIAD-HD.” Real-IAD is introduced instead as a large-scale, real-world, multi-view, high-resolution industrial anomaly detection dataset, with released image resolutions in the 2000–5000 range and native camera resolution $3648\times 5472$ (Wang et al., 2024). The later HiAD paper explicitly curates a high-resolution benchmark from that source dataset, and calls this benchmark RealIAD-HD; the concrete processed dataset used in experiments is RealIAD-2K (Zhang et al., 18 Aug 2025).

A separate later usage appears in the Real-IAD-MVN line. There, the explanatory text states that if someone refers to “RealIAD-HD,” they are essentially pointing at the high-precision, multi-view pseudo-3D variant of Real-IAD, namely Real-IAD-MVN (Zhu et al., 8 May 2026). This suggests that the label is context-dependent: in one context it denotes a 2D high-resolution benchmark configuration, and in another it functions as shorthand for a high-fidelity geometric extension.

2. Source dataset: Real-IAD

RealIAD-HD is derived from Real-IAD, which was proposed to address the saturation and limited scale of earlier industrial anomaly detection datasets. Real-IAD contains 151,050 images of 30 distinct industrial products, with 99,721 normal images and 51,329 anomalous images, and uses a 5 views per sample acquisition protocol consisting of one top view + four side views at 45° (Wang et al., 2024). The dataset covers metal, plastic, wood, ceramics, and mixed materials, and includes both simple parts and more complex assemblies.

Real-IAD also formalizes a broader anomaly taxonomy and evaluation regime than earlier single-view RGB datasets. The paper lists 8 canonical defect types—Pit, Deformation, Abrasion, Scratch, Damage, Missing part, Foreign object, and Contamination—and introduces both conventional unsupervised industrial anomaly detection and Fully Unsupervised Industrial Anomaly Detection (FUIAD) (Wang et al., 2024). It includes image-level labels, pixel-level masks, and sample-level grouping across the five views.

The high-resolution character of Real-IAD is foundational for RealIAD-HD. The paper states that the dataset deliberately increases resolution to capture more subtle defect features, with released images in the 2–5K pixel range after cropping from original captures. This is the direct substrate from which the later high-resolution benchmark configuration is constructed (Wang et al., 2024).

3. Construction and data characteristics of RealIAD-HD

RealIAD-HD is constructed by selecting three representative product categories from Real-IAD: Bottle Cap, Mint, and USB Adaptor. These categories are chosen because they have the smallest average defect areas in the original Real-IAD dataset and are very challenging at lower resolutions (Zhang et al., 18 Aug 2025). To standardize the benchmark, the construction uses only top-view images and uniformly resizes all images to $2048\times 2048$. The resulting processed subset is denoted RealIAD-2K.

Category	Split counts	Total
Bottle Cap	370 train; 299 test normal; 81 test anomaly	750
Mint	306 train; 99 test normal; 322 test anomaly	727
USB Adaptor	362 train; 80 test normal; 102 test anomaly	544
Total	1038 train; 478 test normal; 505 test anomaly	2021

The protocol follows standard unsupervised anomaly detection: training uses only normal images, while testing includes both normal and anomalous images. The anomalies are real defects from the manufacturing process rather than synthetic perturbations (Zhang et al., 18 Aug 2025). RealIAD-HD reuses the pixel-level ground-truth masks from Real-IAD, so evaluation is defined at both image and dense segmentation levels.

The benchmark is explicitly centered on subtle anomalies. The paper states that RealIAD-HD has the smallest relative defect area among the compared high-resolution benchmarks, and that in several examples the relative defect area is less than $0.01\%$ (Zhang et al., 18 Aug 2025). This makes the dataset particularly sensitive to the loss of fine-grained discriminative information under conventional downsampling. A plausible implication is that RealIAD-HD is less a generic high-resolution benchmark than a stress test for methods whose success depends on preserving weak, localized anomaly cues.

4. Evaluation protocol and empirical profile

In the HiAD experiments, RealIAD-HD is evaluated with image-level AUROC (I-AUC), pixel-level AUROC (P-AUC), pixel-level Average Precision (P-AP), pixel-level F1 (P-F1), and Per-Region Overlap (PRO) (Zhang et al., 18 Aug 2025). The image-level anomaly score is defined as the maximum value over the anomaly map $\mathcal{S}$,

$s(X) = \max_{a,b} \mathcal{S}(a,b),$

and pixel-level metrics are computed from $\mathcal{S}(a,b)$ against dense binary ground truth.

Because RealIAD-HD images are $2\text{K}$, the anomaly map and binary mask are downsampled to $512\times 512$ for metric computation. The paper reports that for 2K images, this introduces <0.1% absolute error on all P-metrics (Zhang et al., 18 Aug 2025). The reported evaluation is therefore computationally tractable while remaining close to full-resolution scoring.

Overall results on RealIAD-2K show very high pixel-level separability but substantial variation in segmentation quality across detectors. With HiAD, the reported overall values are: PatchCore with I-AUC 98.37, P-AUC 99.89, P-AP 45.16, P-F1 47.34, PRO 99.10; ViTAD with 94.08, 99.79, 36.59, 41.83, 98.02; DeSTSeg with 93.50, 99.46, 54.67, 53.96, 95.55; RealNet with 93.02, 99.63, 46.84, 48.86, 92.71; FastFlow with 93.66, 99.32, 25.97, 32.82, 96.79; RD++ with 88.03, 98.87, 29.03, 36.27, 94.02; and PaDiM with 89.67, 99.60, 20.74, 28.22, 97.98 (Zhang et al., 18 Aug 2025).

The category-wise results indicate different difficulty regimes. Bottle Cap is described as comparatively easier, with extremely high P-AUC and PRO; Mint is more challenging, with lower I-AUC and P-AP; USB Adaptor combines complex structure with strong overall detectability under HiAD (Zhang et al., 18 Aug 2025). This distribution suggests that RealIAD-HD tests both fine-grained localization and robustness to category-dependent appearance variation.

5. Methodological role in high-resolution anomaly detection

RealIAD-HD functions as a benchmark for methods that must preserve subtle anomalies at native or near-native resolution. In the HiAD framework, the central argument is that conventional low-resolution preprocessing can erase the discriminative evidence required for these defects. The paper makes this explicit through PatchCore comparisons on RealIAD-2K: without HiAD, at $512\times 512$, PatchCore reports I-AUC 91.44, P-AUC 99.51, P-AP 25.53, P-F1 33.43, PRO 95.09; at $3648\times 5472$0, it reports 84.58, 97.71, 10.10, 18.22, 83.98; with HiAD at 2K, the same detector reaches 98.37, 99.89, 45.16, 47.34, 99.10 (Zhang et al., 18 Aug 2025). The concrete implication is that RealIAD-HD is not merely a larger-input benchmark: it measures the extent to which a method can avoid the failure mode induced by downsampling.

HiAD addresses the benchmark with a dual-branch architecture. The high-resolution branch tiles the input image $3648\times 5472$1 into patches and processes them through a detector pool, yielding

$3648\times 5472$2

The low-resolution branch downsamples the image,

$3648\times 5472$3

applies a backbone and detector, and upsamples the resulting anomaly map,

$3648\times 5472$4

The final anomaly map is fused by a pixelwise maximum,

$3648\times 5472$5

The framework further introduces multi-resolution feature fusion,

$3648\times 5472$6

followed by patch features

$3648\times 5472$7

For RealIAD-2K, the paper uses downsampling rates $3648\times 5472$8, equal weights by default, $3648\times 5472$9 detectors, and typically $2048\times 2048$0 high-resolution patches for CNN backbones with stride equal to patch size (Zhang et al., 18 Aug 2025). Most methods use Spatial Clustering Assignment (SCA) on RealIAD-HD, while Retrieval Assignment (RA) is used in the multi-class setting. The benchmark therefore serves simultaneously as a dataset and as a methodological proving ground for patch-based, multi-resolution, detector-assigned inference under constrained compute.

6. Relation to Real-IAD-MVN and the broader “high-fidelity” line

A later development in the Real-IAD family complicates the interpretation of the label “RealIAD-HD.” Real-IAD-MVN is introduced as a high-fidelity geometric extension of the earlier Real-IAD D³ dataset. It replaces sparse 3D point clouds with Multi-View RGB (MV-RGB) + multi-view surface normal maps (MV-NV), using five viewpoints and upgrading geometric resolution to around 0.002 mm normal precision (Zhu et al., 8 May 2026). It keeps the same 20 categories, 69 defect types, and 8,450 samples as Real-IAD D³, but increases the data volume per sample by 5× because each sample is captured from five views.

The Real-IAD-MVN material explicitly states that, in the sense in which people often say “HD” or “high-fidelity,” Real-IAD-MVN is exactly that, and that if someone refers to “RealIAD-HD,” they are essentially pointing at this high-precision, multi-view pseudo-3D variant of Real-IAD (Zhu et al., 8 May 2026). This usage differs materially from the HiAD benchmark usage. In the former, “HD” refers to high-resolution 2D imagery standardized as RealIAD-2K; in the latter, it refers to high-fidelity geometric sensing via dense multi-view photometric-stereo normals.

This suggests that the term “RealIAD-HD” should be interpreted contextually. In high-resolution anomaly detection, it designates a benchmark configuration derived from Real-IAD and centered on three top-view 2K categories. In the high-fidelity geometric branch of the Real-IAD line, it serves as an informal descriptor for Real-IAD-MVN’s dense, multi-view pseudo-3D representation. The common thread is not a single canonical dataset name, but an emphasis on preserving information that is lost in lower-fidelity pipelines: either fine image detail in RealIAD-2K or dense micro-geometry in Real-IAD-MVN.