Research report: Beyond hue labels: a data-driven analysis of lightness–chroma variation in online fashion assortments

When an online shopper searches for a “red knit,” the results that appear are rarely one kind of red. They range from dark wine tones to bright cherry reds, muted brick shades, and highly saturated vivid reds, yet all of them sit under the same categorical label. A new study published in the journal Fashion and Textiles argues that this familiar way of organizing products conceals a large share of the color information that actually matters, and it offers a quantitative framework for measuring what hue labels leave out.

The research, led by Subeen Son, Eung Tae Kim, and Juhyun Ro, analyzed product images from MUSINSA, a major South Korean online fashion retail platform. On February 9, 2026, the team collected the first 200 products returned for each of four color–garment keyword queries equivalent to “red knit,” “green knit,” “blue knit,” and “yellow knit,” using the platform’s default recommendation-based ranking. The resulting dataset contained 800 images, 200 per hue query. The researchers deliberately did not impose any independent color criterion to reclassify or exclude products, because their goal was to characterize exactly the assortments that users encounter under these queries, not to build a validated color taxonomy.

Each image was processed with a pipeline designed to isolate the garment itself from the visual noise of online retail photography, which includes varied lighting, backgrounds, styling contexts, and shooting conditions. A pretrained semantic segmentation model based on the SegFormer-B4 architecture was used to retain only pixels classified as the upper-body garment, excluding backgrounds and other regions. Within each isolated garment region, MiniBatch K-means clustering was applied to pixel color values in the CIELAB color space, a perceptually oriented system in which equal numerical distances correspond more closely to equal perceived differences than in device-dependent RGB. The dominant garment color was defined as the centroid of the cluster containing the largest share of pixels, and from that color the researchers computed two continuous coordinates: lightness (L*) and chroma (C*), the latter measuring the distance of the color from the neutral axis in the a*–b* plane.

The central finding is that products retrieved under the same hue-based query occupy broad and differently configured regions of the L*–C* plane rather than clustering tightly around a single tone. In the red assortment, the observations formed a broad diagonal distribution extending from darker, less chromatic regions toward lighter, more saturated ones. The green group showed wide variation in both dimensions, including products with similar lightness but markedly different chroma. The blue group likewise spread across a wide region of the plane, while the yellow assortment concentrated at relatively high lightness values but still varied substantially in chroma. In other words, the hue term itself conveyed only a partial description of the color composition of each product set.

To summarize these continuous distributions without imposing predefined tone categories, the researchers applied K-means clustering to the product-level L*–C* coordinates and selected the number of clusters adaptively. Candidate values from 2 to 10 were evaluated using three internal validity indices—the Silhouette coefficient, the Calinski–Harabasz index, and the Davies–Bouldin index—ranked separately for each index, with the ranks summed to choose the final partition. Because the three indices often favor different solutions, no single candidate was consistently optimal. For the red assortment, the Silhouette coefficient peaked at k = 2, the Calinski–Harabasz index at k = 9, and the Davies–Bouldin index at k = 5; the rank-aggregation procedure selected k = 5. Applied independently across hues, the procedure yielded five partitions for red, four for green, four for blue, and three for yellow, each summarizing where products concentrated within that assortment’s distribution.

The cross-hue comparison revealed that these configurations were not interchangeable. In the red assortment, cluster mean lightness ranged from about 17.6 to 46.0 and mean chroma from about 32.1 to 68.4, with two clusters at nearly identical mean lightness (44.7 and 46.0) but very different mean chroma (68.4 versus 34.4). The green assortment spanned mean lightness from 19.3 to 73.8 and mean chroma from 9.3 to 39.0, showing that low-chroma products occurred across widely different lightness levels. In the blue assortment, two clusters with similar mean lightness (27.5 and 24.7) differed considerably in mean chroma (42.6 versus 18.5). The yellow assortment, with only three clusters, concentrated products at high lightness, where two clusters with mean L* of 82.4 and 79.5 differed substantially in mean chroma (21.9 versus 47.7). Principal component analysis added a geometric view: the first principal component explained 83.1% of the variance in red, 81.2% in blue, and 77.3% in green, but only 53.7% in yellow, and its orientation differed across hues, from 12.7° for yellow to 156.3° for blue.

The team then tested whether the framework could track assortment change over time. Using the identical search and sampling procedure on August 8, 2026, they collected a second set of 800 images and compared the two collection periods. The results were strongly hue-specific. The green assortment showed the largest centroid shift in L*–C* space (6.842), driven mainly by a decrease in mean lightness from 47.5 to 40.6, along with a substantial drop in mean dispersion from 21.9 to 15.6; its February and August distributions differed significantly in a permutation test using the energy distance statistic (p = .001). Yellow shifted primarily along the chroma dimension, with mean C* falling from 29.0 to 24.6 while lightness stayed nearly stable (p = .003). Red showed only a small centroid shift of 2.016, yet its overall distribution differed significantly (p = .017), indicating that distributional change was not captured by central location alone. Blue, by contrast, remained comparatively stable, with a centroid shift of 2.077, nearly unchanged dispersion, and no statistically significant distributional difference (p = .549). Forty-four products appeared at both dates, and the authors note that the permutation test did not account for this potential dependence.

The authors are careful about what these results do and do not establish. The data-adaptive partitions are descriptive summaries of empirical distributions, not perceptually validated tone categories, and K-means necessarily partitions continuous observations regardless of whether naturally discrete classes exist. The retrieved product sets reflect the assortments presented by a single platform under its proprietary search and ranking system at two specific dates, so the findings should not be read as population-level distributions of red, green, blue, or yellow fashion products. The temporal analysis involved only two collection dates and cannot establish seasonal patterns or systematic trends; the observed differences are interpreted as period-specific assortment changes. The analysis also focused on knit products and four broad hue categories, leaving generalization to other garment types, retailers, and markets to future work.

Within those limits, the study offers a practical contribution to digital fashion retail. Mapping products in L*–C* space gives retailers, merchandisers, and platform operators a quantitative view of how a hue-specific assortment is distributed across lightness and chroma, and the adaptive partitions summarize where products concentrate. Designers could use such distributions to identify densely and sparsely represented regions when planning future color options, and the same image-processing pipeline could in principle be integrated with existing hue metadata to let consumers further order, group, or filter products within a broad color category after selecting a hue. The authors stress that these consumer-facing applications remain prospective: the study did not implement a retrieval or recommendation system, nor did it test whether users prefer lightness–chroma-based filtering or whether it improves search efficiency, satisfaction, or purchasing outcomes. Its demonstrated contribution is a measurement and monitoring framework, with effectiveness of extensions left to future system implementation, user studies, and retail experiments.

The broader message resonates beyond e-commerce. Consumers rarely describe garments with hue terms alone, instead reaching for modifiers such as light, vivid, muted, or dark, and prior research has shown that inaccurate color representation in online retail can lead to dissatisfaction, increased returns, and customer defection. By making within-hue variation measurable across assortments and collection dates, the framework connects image-derived color characteristics with assortment-level analysis in a way that nominal labels alone cannot. Its current value, the authors conclude, is analytical: the distributions and partitions describe the products actually observed in the sampled assortments, they do not define universal tone categories or optimal color structures, and further evaluation across platforms, product categories, and consumer contexts is needed before the approach reaches consumer-facing product organization, retrieval, or recommendation.

Subject of Research: Data-driven analysis of lightness and chroma variation within hue-labeled online fashion assortments

Article Title: Beyond hue labels: a data-driven analysis of lightness–chroma variation in online fashion assortments

Article References: Son, S., Kim, E. T., & Ro, J. (2026). Beyond hue labels: a data-driven analysis of lightness–chroma variation in online fashion assortments. Fashion and Textiles, 13(1), Article 37. https://doi.org/10.1186/s40691-026-00495-6

Image Credits: AI Generated

DOI: 10.1186/s40691-026-00495-6

Keywords: CIELAB, lightness, chroma, online fashion retail, color analysis, K-means clustering, semantic segmentation, principal component analysis, assortment monitoring, e-commerce, MUSINSA, color planning