K-Fashion is a Korean fashion dataset built in 2020 by OpinionLive, together with Ewha Womans University, the Korea Fashion Industry Research Institute, AI.M and Wearly, and released on AI Hub. The full set runs to roughly 1.2 million images, annotated not just with a style label but with polygons, bounding boxes and attributes for each fashion item. With 66,321 views and 4,059 downloads, it is a heavily used asset in Korean fashion-AI research.

This diagnosis used the 967,806 training images. Below is a collage that captures the dataset's overall mood at a glance. In words: most of them are lookbook-style photos of a full standing body shot in a bright indoor setting. That consistent shooting format becomes an important foreshadowing for the analysis to come.

On the surface it's an abundant dataset. The problem is that the photos aren't spread evenly across the 24 styles. Street alone holds 449,494 images (46.5%), nearly half the total, and more than five times the runner-up, Feminine (88,325). At the other end sit tiny classes of roughly a thousand images or fewer: Punk at 382, Hip-hop at 1,240, Western at 1,712. The average is 40,325 images per class, but the standard deviation is 91,811, larger than the mean. In other words, the distribution doesn't cluster around the average; it splits to the extremes in a long tail.

Level 1 is a pixel-level fitness check. The short version: the hygiene items are fine. Every image is uniformly RGB, no missing values turn up, and label integrity shows nothing unusual. One thing does stand out: the spread in resolution. The smallest image is 280×385px and the largest is 4160×3872px, a wide size gap within a single dataset. That's a point where you'll want a clear resize policy before training.

The interesting part is the "mean image." Below is the pixel-wise average of all 967,806 photos. What survives is a hazy human silhouette with almost no defined shape. The blurrier the mean, the more the underlying photos differ from one another — so we can read it as a signal of high visual diversity across the dataset.

We can build a mean image per class, too. For the six classes DataClinic provides (Other, Modern, Sexy, Kitsch, Punk, Hippie), we placed each class's actual representative sample (left) next to its class mean image (right). The sharper the mean image, the more consistent that class's framing and pose.

Level 2 looks at every image once through a general-purpose neural network called Wolfram ImageIdentify Net V2 (observation dimension 1,280). Because it's a general recognition model, not one specialized for fashion, it shows "how this dataset scatters when viewed through generic visual features." The key metric here is density. A high-density image is a "typical" one, surrounded by many similar images; a low-density image is an "outlier," with few look-alikes nearby.

First, the overall density distribution. Most images cluster tightly around 0.1, with a thin tail trailing right toward 0.3. It's a single mass, with a small set of typical images claiming the higher densities.

Split by class, differences appear. In the box chart below, the leftmost class's box is lower and shorter than the others. That class is "Other," a catch-all for photos that don't fit a defined style. With few look-alikes among them, its density comes out lowest. The remaining classes have similar medians of 0.12–0.13, but their upper whiskers stretch all the way to 0.6–0.7, meaning each class contains a separate cluster of very typical photos.

The PCA map below lays that distribution flat. The 960,000 points spread out into something close to an asymmetric triangle. The gray diamond (the overall mean feature) sits off to one side rather than in the center, which we can read as the result of Street's overwhelming volume pulling the distribution's center of mass toward itself.

The same data as a density heatmap is relatively pale inside, with one dark hotspot rising at the top edge. That's a signal that K-Fashion doesn't gather around a single center but forms a separate cluster of typical images out at the edge.

Per-class density curves with representative images make it more concrete. Modern records the highest density (top-density sample around 0.70), its lookbook-style photos packed tightly together; Hippie and Kitsch also form fairly distinct peaks. Other, by contrast, has a low, wide peak and weak cohesion.

Level 3 keeps the same ImageIdentify Net V2 as its base but applies a lens optimized down to 150 dimensions to preserve class-discriminative power. It's the stage that sharpens the between-class differences the general lens missed. One caution, though: L3's density values are computed in a different dimensional space than L2's, so you can't compare the two stages' numbers directly. The only thing you can compare is the "shape" of the distribution.

The L3 density histogram peaks around 0.55, and its right tail reaches far further than L2's (beyond 1.5). That means a wider dynamic range — a handful of hyper-typical images claim extremely high densities.

In the box chart too, class medians become more uniform than L2's, at 0.50–0.65. Here again the leftmost Other is the lowest. Several classes' upper whiskers cross 2.0, showing that each class holds extremely typical photos of its own.

The most striking change is the shape of the PCA map. L2's asymmetric triangle tidies into a near-symmetric ellipse in L3. We can read this as the 150-dimensional optimization correcting, to some degree, the directional skew that Street's volume had created. Place the two maps side by side and the difference is clear.

The L3 density heatmap differs again from L2's. It's a donut shape — empty in the middle, with two dark hotspots rising at the edges (especially left and right). That's visual evidence that K-Fashion exists not as one center but as several archetypal clusters.

L3 brings into focus minority classes that weren't prominent in L2. Sporty (14,701 images) forms the hottest distribution, and Hip-hop (1,240) starts to separate into its own cluster. Sexy's top-density sample climbs to about 1.44, showing strong typicality. The point: even a class with few images can form a clear peak under the class-discriminative lens, as long as its visual consistency is high.

This is the heart of the diagnosis. In the L3 class-discriminative space, the #1 and #2 "most typical" (highest-density) images out of all 960,000 both belong to the Street class, at densities 2.21 and 2.15. And it isn't only the top two: 9 of the top 10 highest-density images are Street, their densities packed into a narrow band from 2.09 to 2.21. It seems natural for a class with overwhelming volume to occupy the most-typical slots. But open those top images and the story changes.

The #1 image is a full-body shot against a white studio backdrop: a floral chiffon blouse, blue skinny jeans, red flats. Visually, anyone would call it Romantic or Feminine. And #1 and #2 are consecutive frames from the same shoot, with only a slight change in composition. Two nearly identical photos split first and second place for "most typical Street."

This happens because of batch shooting. A lookbook shoot has one model wear one outfit and fire off dozens of frames on the same set. The result is a pile of nearly identical photos (near-duplicates) within a single class. The more similar photos there are, the higher the density around them, and the higher the density, the more the model learns that pattern as "typical." When 490,000 images meet batch shooting, what the model learns isn't a style. It's a shooting format: "studio + full standing body + bright top + denim."

Look at the opposite end and the message sharpens. The lowest-density image in the same Street class (density 0.30) is a shot of a black oversized graphic sweatshirt, perched on a city windowsill. Real buildings and a factory complex are visible through the window. Anyone would call it genuine street feeling. Yet this photo is classified as "atypical" and pushed to the margins of the distribution.

K-Fashion is a valuable asset — an overwhelming 970,000 images with detailed annotation. It passes the L1 hygiene check cleanly. But DataClinic's distribution lens reveals three structural traits the accuracy number doesn't show. Set out as three comparison frames, they look like this.

The message of this diagnosis comes down to one sentence. A correct label doesn't make the data right. Even when every photo is labeled "Street" exactly, if near-duplicates poured out of a single session dominate the cluster, the model learns the photo set, not the clothes. Data quality isn't just about label accuracy. It's about whether the distribution that label points to matches the concept we intended.

Practically, three directions stand out: reducing batch-shoot near-duplicates or weighting by cluster; downsampling bloated classes like Street while reinforcing minority classes like Punk and Hip-hop; and revisiting wobbly label boundaries like Sophisticated ↔ Feminine. Which to do first depends on whether the service you build with this data is search, recommendation, or trend analysis.

You can find the full diagnostic process and figures in DataClinic Report #127, and the original data on the AI Hub K-Fashion page. If you want to explore the data's cluster structure yourself, we recommend unfolding the embedding with Pebbloscope.