HabinKim/kfashion-visual-distinctiveness

Is "K-Fashion" Visually Distinct? — Brand Selection & Analysis Procedure

'K-패션'은 시각적으로 구별되는가 — 브랜드 선정 및 분석 절차 공개

A reproducible release for a study asking whether Korean fashion brands are visually distinguishable from global brands under a fashion-specialized vision model (Fashion-CLIP). Original product images are NOT redistributed (copyright); we share derived embeddings, metadata, the brand-selection record, results, and all code so the analysis is fully reproducible.

한국 브랜드 옷이 글로벌 브랜드 옷과 시각적으로 구별되는가를 패션 특화 AI(Fashion-CLIP)로 검증한 연구의 재현용 공개 자료입니다. 원본 이미지는 저작권상 비공개, 대신 임베딩·메타데이터· 브랜드 선정기록·결과·코드 전체를 공개합니다.

TL;DR — Findings

• —Korean vs global brands are not visually separable: origin silhouette ≈ 0.02 (0 = fully mixed, 1 = fully separated); not statistically significant under a brand-level permutation test, with or without photo-format control, at every price tier.
• —Clothes cluster by individual brand, not by nation.
• —A first, dramatic "Korean brands don't use models" signal turned out to be a sampling artifact (4-brand sample); it vanished with a properly selected 18-brand sample (both origins ~73% flat-lay).

Research design

• —Origin × price-tier grid: KR / GLOBAL × {SPA, Contemporary, Designer}, 3 brands per cell → 18 brands.
• —~600 product images per brand from Naver Shopping search API, 6 garment categories each (balanced).
• —Total: 10,640 images (KR 5,236 / GLOBAL 5,404).

Brands (18)

Brand-selection procedure (LLM-assisted, reproducible)

The brand set was chosen with a transparent, bias-reducing pipeline (recorded in data/brand_selection_candidates.csv and brand_selection_final.csv):

1. 1.Criteria fixed first (before any LLM): origin definition; price-tier strata (to avoid confounding origin with segment); single brand (no multi-brand retailers); must be collectable on Naver (≥150 products).
2. 2.Independent LLM proposers (×3) with different perspectives — market share, editorial/critical, online/Gen-Z — each proposed 4–5 brands per cell.
3. 3.Consensus by vote count (brands proposed by multiple agents preferred).
4. 4.Verification: factual existence / origin / single-brand checks, and collectability via Naver API product counts; ambiguous or low-yield brands dropped.
5. 5.Finalize balanced 18, equal N per brand; full record kept for the appendix.

Analysis pipeline

1. 1.Embed each image with Fashion-CLIP (patrickjohncyh/fashion-clip, 512-d, L2-normalized, cosine).
2. 2.Photo-format control: zero-shot classify model vs flat-lay product shots.
3. 3.Origin separability: silhouette (cosine), nearest-neighbor cross-origin rate, centroid distance.
4. 4.Segment control: repeat within each price tier.
5. 5.Significance: bootstrap 95% CI + brand-level permutation test (shuffles which brands are KR/GLOBAL — the honest unit, since images within a brand are not independent).

Files

• —data/brand_embeddings.parquet — 10,640 × (brand, origin, tier, shot_type, e0…e511). Reproduce all analysis from this; no images needed.
• —data/brand_product_metadata.csv — per-product metadata (image URL, category, price, mall, productId). Control variables; images not included.
• —data/brand_selection_candidates.csv / brand_selection_final.csv — selection record (votes, Naver counts).
• —results/*.csv — format-control, tier-control, significance summaries.
• —figures/*.png — column/paper figures.
• —code/*.py — collection + analysis scripts (Naver API key NOT included).

Reproduce

import pandas as pd, numpy as np
from sklearn.metrics import silhouette_score
df = pd.read_parquet("data/brand_embeddings.parquet")
X = df[[f"e{i}" for i in range(512)]].values
print(silhouette_score(X, df["origin"], metric="cosine"))  # ≈ 0.02

Full pipeline: see code/ (collect → crop → embed/classify → tier → significance).

License & ethics

• —Code: MIT. Derived data (embeddings, results, metadata): CC-BY-4.0.
• —Original images are NOT included/redistributed (Naver listing copyright); only image URLs and derived 512-d vectors are shared. Data collected via the official Naver Shopping Search API for non-commercial research.
• —Model: Fashion-CLIP (patrickjohncyh/fashion-clip).

Limitations

• —Single embedding model (Fashion-CLIP); cross-model + human-rater validation recommended.
• —Images from a marketplace (resellers), not official brand channels.
• —The "Hallyu halo / spillover" reading is an interpretation/hypothesis, not tested here.

Sources & attribution (출처 표기)

• —Model — Fashion-CLIP: Chia, P. J., et al. (2022). "Contrastive language and vision learning of general fashion concepts." Scientific Reports, 12, 18958. Model weights: `patrickjohncyh/fashion-clip`. Used as-is (no fine-tuning by us).
• —Data, brand study (this release): collected via the Naver Shopping Search Open API (search.naver.com), June 2026, for non-commercial research. Only image URLs + derived embeddings are shared; original images are not redistributed.
• —Data, style study (companion, NOT in this release): AI-Hub 'K-Fashion 이미지' dataset, built with support of the Ministry of Science and ICT & NIA, Korea (aihub.or.kr). 본 연구는 과학기술정보통신부·한국지능정보사회진흥원(NIA)의 지원으로 구축된 'K-Fashion 이미지' 데이터(AI-Hub)를 활용하였다.

Citation

Kim, Habin (2026). Is "K-Fashion" Visually Distinct? Brand Selection & Analysis.
Hugging Face dataset. (column: 「김하빈의 스몰데이터」)

Model: Chia, P.J., et al. (2022). Contrastive language and vision learning of
general fashion concepts. Scientific Reports, 12, 18958.