Fashion Product Images Dataset (44k) — Kaggle

It is the reference apparel catalogue corpus in machine learning: 44,000+ products from Myntra, a major Indian fashion e-commerce platform, each described by manually-entered attributes and photographed professionally at 2400x1600. One master attribute table carries all ten columns per SKU; each row's id locates that product's photograph inside the image folder and a JSON record holding descriptive text about the item.

This is the full-resolution edition, packaged at roughly 24.8 GB because it stores every photograph at original size. A schema-identical lightweight sibling resizes the same 44,000 photos down to roughly 565 MB for prototyping, and because the attribute schema matches column for column, anything trained there transfers upward unchanged. The publisher's own suggested starting points — image classification from the top category column, NLP over the per-product descriptions to predict categories, extension to the remaining label attributes — all assume exactly this join structure.

Engagement says the niche is real: roughly 459,000 views, over 108,000 downloads and 274 public notebooks, a perfect 1.0 usability rating, and a slot among the Kaggle-published records in Datadory's 1,744-dataset catalog. Within our apparel, accessories & luxury goods slice it is one of only three records carrying native image files alongside structured rows.

One line per SKU, every label attached:

id=15970 gender=Men      masterCategory=Apparel     subCategory=Topwear    articleType=Shirts
         baseColour=Navy Blue season=Fall year=2011 usage=Casual
         productDisplayName="Turtle Check Men Navy Blue Shirt"
         image=images/15970.jpg      text=styles/15970.json

id=39386 gender=Men      masterCategory=Apparel     subCategory=Bottomwear articleType=Jeans
         baseColour=Blue      season=Summer year=2012 usage=Casual
         productDisplayName="Peter England Men Party Blue Jeans"
         image=images/39386.jpg      text=styles/39386.json

id=59263 gender=Women    masterCategory=Accessories subCategory=Watches    articleType=Watches
         baseColour=Silver    season=Winter year=2016 usage=Casual
         productDisplayName="Titan Women Silver Watch"
         image=images/59263.jpg      text=styles/59263.json

Three facts worth noticing. The category columns form a strict hierarchy — Apparel contains Topwear contains Shirts — so a prediction at any level can be checked against the other two. The display name repeats the colour word, which makes baseColour verifiable from free text. And both artefacts fall out of the id mechanically: pixels at images/<id>.jpg, prose at styles/<id>.json, so joining photographs and descriptions to labels needs no lookup table.

Ten columns, each verified against the published schema rather than inferred. Seven of them are closed sets: 5 gender segments, 7 top-level categories, 45 subcategories, 141 article types, 46 colours, 4 seasons and 8 usage contexts. That enumerability is what separates this table from hand-assembled catalogues where every attribute arrives as unstructured text needing normalisation before a model can touch it.

Beyond the table proper sits the corpus's quiet advantage: a per-product JSON record holding descriptive text about each item's characteristics. It lives outside the ten-column schema, so it ships as an additional field on request alongside brand parsed out of the display name and two-level category rollups.

Geography: India — the entire corpus comes from one marketplace's catalogue, so merchandising conventions are consistent end to end: season tags tuned to Indian retail cycles, colour naming in marketplace style, and a gender taxonomy that includes kids' segments.

Temporal: a static snapshot. Product years run roughly 2009–2017, the record was last touched in March 2019, and its expected update frequency is set to never. That frozen quality is a feature when you need reproducibility — 181 of the 1,744 datasets in Datadory's catalog never update, and research corpora cluster exactly here — but recency has to come from a living complement such as the Lyst catalog or the Zalando catalog.

Granularity: one row per SKU with one corresponding photograph and one descriptive-text record. No transaction history, no prices, no stock levels — the schema stops at description, because the corpus was assembled for computer vision rather than commerce analytics.

API, files, or your warehouse. Daily, weekly, or hourly.

Pick the cadence and the landing zone; the same ten-column rows arrive either way, with full-resolution photographs and descriptive-text records referenced by id so your storage layer decides whether they travel inline or stay object-addressed. At 24.8 GB the corpus rewards a warehouse landing zone, and the sample comes first — rows, field dictionary and image set confirmed against your use case before any commitment.

• Image classification at production weight — train over the 141 article types on 2400x1600 photography, then evaluate whether predicted labels respect the subCategory constraints already encoded in the table.
• Attribute extraction with text supervision — the display name embeds colour words beside a curated colour vocabulary, and the per-product descriptions add a second weak-supervision channel for tagging models.
• Visual search and retrieval — embedding pipelines get real storefront imagery at full resolution, with category labels ready-made for evaluating nearest-neighbour quality.
• Taxonomy benchmarking — 7 master categories fanning into 45 subcategories is a compact reference hierarchy for teams defining their own product-data intake standards.
• Assortment research — gender segment, season and usage context slice the catalogue into readable merchandising cuts without any modelling.

Personas ranked by fit: data scientists and developers lead; competitive intel product teams, market researchers and e-commerce operators follow for taxonomy and assortment work.

Within its own family this record is the completeness anchor. The lightweight sibling carries identical rows and schema at prototype weight; the ashraq Hub mirror reshapes the catalogue as Parquet with images embedded in-row, while the ceyda copy trims to 42,700 rows and three label columns. This edition keeps everything: all 44,000 products, the full ten-column schema, original-resolution photography and the descriptive-text layer none of the reduced copies carry.

Against the wider vision corpora the trade runs the other way: DeepFashion brings 800,000+ annotated images under research-use restrictions, and Fashionpedia pairs 46,781 images with 342,182 segmentation masks. Those are annotation corpora; this is a storefront corpus. If you want what shoppers actually saw — 44,000 labelled retail SKUs photographed the way marketplaces shoot them — that is this record's lane, and the Apparel Images Dataset is its nearest neighbour in kind.

Datadory scores it 9 out of 10 on its rubric of field documentation, access reliability and freshness — above the catalog-wide mean of 7.81, and the top of the fashion-corpus family in this industry slice. Field definitions are marked verified against the published schema, the schema has been stable since publication, and engagement metrics double as a health signal: nearly half a million views and six-figure downloads make it among the most-exercised corpora we track.

Two caveats belong in the open. First, rights over the underlying imagery trace to the marketplace the corpus was drawn from, and that chain is unverified against the rights holder — treating rights verification as part of procurement is standard practice here. Second, whether the per-product descriptive-text records carry pricing detail or only characteristic prose stayed unstated during the August 2026 research pass; we pin it against your sample rather than assert it.