Executive Summary

The "Korean Traditional Ink and Color Painting Production Data" is a public dataset released through AI Hub. It was built to digitize paintings created with traditional Korean techniques — ink, brush, and watercolor pigments — and make them available for AI training. The dataset permits commercial use and can be diagnosed on DataClinic as Report #194.

▲ 3,995 "Korean Traditional Ink Paintings" — Bentleys and tanks nestled among tigers and chrysanthemums (DataClinic L1 collage)

The gap between expectation and reality hits you the moment you open the collage. Familiar traditional subjects are all present — landscape paintings, tigers, carp, chrysanthemums. But right next to them sit luxury sedans, fighter jets, and smartphones. Every image follows the same composition: a single subject centered on a white background. The technique is unified — watercolor and ink strokes throughout — but the subjects ignore centuries of history.

Hear the words "traditional Korean ink painting" and certain images come to mind. The tiger as mountain guardian, the carp leaping through the Dragon Gate, the chrysanthemum representing the recluse among the Four Gentlemen, and the pine tree embodying integrity. These traditional subjects — called hwaje (畫題) — have anchored Korean painting for centuries.

But open this dataset and those traditional subjects share shelf space with cars (79 images), motorcycles (78), phones (75), airplanes (72), tanks (66), and blenders (57). Below, we place six traditional subjects side by side with six modern objects. In each card, the left image is an actual sample and the right is the class mean image.

Traditional Subjects — The Heart of Korean Painting

Actual

Mean

Tiger (43 images)

Actual

Mean

Carp (32 images)

Actual

Mean

Chrysanthemum (36 images)

Actual

Mean

Pine Tree (45 images)

Actual

Mean

Bear (27 images)

Actual

Mean

Mountain (43 images)

▲ Six traditional subjects — averaging 37.7 images/class. Left: actual sample / Right: mean image

Modern Objects — The 21st Century Enters "Traditional Painting"

Actual

Mean

Car (79 images)

Actual

Mean

Airplane (72 images)

Actual

Mean

Tank (66 images)

Actual

Mean

Phone (75 images)

Actual

Mean

Motorcycle (78 images)

Actual

Mean

Blender (57 images)

▲ Six modern objects — averaging 69.5 images/class, nearly twice the traditional count

Level I examines the basic physical properties of images. There are no missing values and channels are consistently RGB. The trouble, however, shows up in image dimensions.

Image Size Consistency: Poor

Resolutions range from 3,040×2,150 px to 9,930×7,020 px — a 3.3× size ratio that far exceeds DataClinic's 1.5× threshold. Since every image is a high-resolution scan (at least ~6.5 megapixels), downsampling to a common size is straightforward, but as-is, a resizing step is mandatory before model training.

What the Overall Mean Image Tells Us

Stack all 3,995 images on top of one another and what do you get? The global mean image is a pale grayish-beige field with a blurry brown mass at center. No structure, no texture, no identifiable form.

▲ Mean of all 3,995 images — confirms every painting shares the same "subject centered on white background" composition

This confirms that every image follows an identical compositional template: a single subject centered on a white background. That is an illustration convention, not the aesthetic of traditional ink painting with its deliberate negative space and compositional asymmetry. When DataClinic flags "low diversity," it means pixel-level visual diversity. Semantically there are 74 different subject types, but the visual patterns AI perceives are remarkably uniform.

Level II observes images through the 1,280-dimensional general-purpose lens of Wolfram ImageIdentify Net V2. Instead of raw pixels, it diagnoses the geometric structure and density distribution of data in a high-dimensional feature space.

PCA — One Cluster, Internal Density Gradient

The PCA chart projecting 1,280 dimensions down to two shows all 3,995 images forming a single large cluster. It is slightly elongated and lean. No clearly separated sub-clusters appear, but density varies across the interior: the center-left region is dense, while points thin out toward the periphery.

▲ L2 PCA — a single cluster with density gradient. Orange dot: feature-space origin. Red dot: PCA-space origin.

Density Heatmap — API Claims vs. Visual Reality

The DataClinic API describes this distribution as "bell-shaped around the origin and data center." The actual density heatmap tells a different story: the distribution is asymmetrically concentrated in one corner. The high-density core (deep red) occupies a limited area, while the rest of the space spreads out at low density. "Asymmetric concentration" is a more accurate description than "bell-shaped."

▲ L2 density heatmap — asymmetrically concentrated distribution, diverging from the API's "bell-shaped" description.

Per-Class Density — High vs. Low

Zoom into individual classes through the Wolfram lens and a clear pattern emerges: simple, symmetrical objects (bowls, bottles) cluster tightly at high density, while irregular organisms (bats, octopuses, rays) scatter widely at low density.

Density

Actual

Bowl/Plate — Density 0.123 (Highest)

Compact high-density cluster. The class AI is most confident about.

Density

Actual

Bottle — Density 0.115

Simple cylindrical form. Minimal pose variation yields high internal consistency.

Density

Actual

Bat — Density 0.075 (Lowest)

Traditional symbol of good fortune. High pose variation and extreme ink-wash contrast scatter the cluster.

Density

Actual

Ray — Density 0.079

Irregular silhouette. Only 28 samples compound the low count with shape variation.

▲ Each card — Left: L2 density chart / Right: representative sample (actual image)

Nearest & Farthest Neighbors — Two Worlds in One Dataset

DataClinic's similarity analysis identifies the nearest and farthest neighbors in feature space. Nearest neighbors cluster by visual shape: bat-dinosaur-butterfly-ray share dark silhouettes, while otter-hamster-rabbit bond over similar small-mammal body plans.

The farthest pair, however, exposes the dataset's fundamental contradiction. The two classes most distant in feature space are bat and airplane.

Bat (蝹蝠 = fortune)
Density 0.075

↔

Airplane
Modern vehicle

▲ The two most distant classes in feature space: the traditional good-fortune bat and the modern airplane. Two fundamentally different worlds coexist in a single dataset.

Level III applies an optimized 48-dimensional lens extracted from a 265-layer neural network trained on 3 million images. Its purpose is to reveal domain-specific structures that the general-purpose 1,280-dimensional lens misses.

PCA — More Compact, but Internal Structure Stands Out

The PCA chart through the 48-dimensional lens shows a more compact, rounder distribution than L2. The optimized lens pulls data points closer together. Yet within that tighter form, density variation is actually more pronounced. Broad groupings — artifacts vs. natural subjects, large animals vs. small objects — emerge as distinct sub-groups.

▲ L3 PCA — more compact than L2, but internal density variation is more pronounced.

Density Heatmap — Bimodal Structure Appears

The L3 density heatmap reveals a bimodal structure invisible in L2. One high-density concentration sits in one region, and a second density nucleus appears at a distance. The 48-dimensional lens is beginning to separate the two lineages within the data — likely the artifact family and the natural-subject family.

▲ L3 density heatmap — bimodal structure more pronounced than L2. Correlation 0.999 between L2 and L3, showing two resolutions of the same phenomenon.

Three Density Zones

Combining L2 and L3 analysis, the 74 classes fall into three density zones.

No discussion of art datasets is complete without WikiArt. Diagnosed as DataClinic Report #115, WikiArt contains 81,471 Western art images classified into 27 styles (Impressionism, Baroque, etc.). Overall score: 53 (Poor) — close to the 57 of the Korean ink painting dataset. Both score "Poor" in the art domain, but the reasons are polar opposites.

WikiArt's difficulty is inherent. Even art historians debate where Monet ends and Renoir begins. Classifying artistic style is an intrinsically hard task for AI — a feature of the domain, not a flaw of the dataset.

Dataset #194's problem is structural. Telling a tiger from a car is trivial — no AI confuses the two. The question is: "Why are they in the same dataset?" Painting a tank in ink-wash technique does not make it traditional Korean art.

A data quality diagnosis does not end at the numbers. The real question is: "What goes wrong when you actually build an AI system with this dataset?" Answering that completes the diagnosis.

What Density Tells Us About Misclassification Risk

The density analysis yields its most paradoxical finding when we compare these two classes.

DataClinic recommends Data Bulk-up for this dataset. Class-level sample counts are imbalanced and overall density is low, so more images are needed. A fair point — but the fundamental problem runs deeper.

Priority 1: Domain Curation

The most urgent step is not adding images but reorganizing the ones already there. Two paths forward:

• A. Separate modern objects: Move cars, airplanes, tanks, smartphones, and other non-traditional subjects into a separate dataset. Reconstruct a pure traditional painting dataset from the remaining natural subjects and traditional architecture.
• B. Rename the dataset: Change "Korean Traditional Ink and Color Painting" to something like "Multi-Subject Ink/Watercolor Technique Dataset." The technique is consistent, so a technique-centered name is more accurate.

Priority 2: Bolster Traditional Subject Classes

Increase sample counts for under-represented traditional subjects — bear (27), carp (32), ray (28) — to 70+ images each, matching the artifact-class average of 68.8.

Priority 3: Pose Control for Low-Density Classes

For sparse classes like bat (0.075), octopus (0.075), and ray (0.079), new images should maintain some consistency in pose and composition to tighten the cluster. Add volume, but within controlled variation so that density improves.

Priority 4: Standardize Image Dimensions

The 3.3× size ratio must be resolved. Since all images are high-resolution scans, downsampling to a common resolution (e.g., 4,000×3,000 px) is practical.