Product Image Dataset for Visual Search & Similarity Engine Training

A product image dataset for visual search built for teams training image similarity models, shop-by-image workflows, and product matching engines across many retailer photography styles.

Home / Use Cases / Visual Search

Used by companies like Syte.ai, Visenze, Snap, Pinterest Lens, and Google Lens, as well as in-house teams at retailers building "shop the look" and "find similar" features as part of an ecommerce visual search dataset strategy.

Problem statement

Visual search models require image similarity training data that spans the same product category across many different visual styles — different colors, materials, backgrounds, and shooting angles.

Training a visual search engine dataset on a single retailer's images produces a model that works well for that retailer and poorly everywhere else.

How ImageHub solves it

  • Same product category (e.g. sofas) represented across 10-20 different retailer photography styles.
  • CLIP-derived semantic tags for fine-grained attribute labels (material, color, style, pattern).
  • Primary image flag distinguishes the main product view — the most important for similarity indexing.
  • Category hierarchy for hard-negative mining (sofa vs chair vs ottoman — close but distinct).

What to request

  • Request cross-retailer bundles for the same category — this is what makes it a genuine product matching image dataset rather than a single-brand set.
  • Use CLIP tags for attribute filtering when building attribute-aware similarity models.
  • Request primary images only for catalog-matching use cases; primary + carousel for multi-angle embedding, ideal for a shop by image training dataset.

Key fields / metadata delivered

  • clip_tags: semantic attributes (color, material, style, pattern)
  • category_path: hierarchical category for hard-negative mining
  • primary_image_flag: TRUE = main product image for similarity indexing
  • site_name: retailer source for domain diversity analysis

Retail AI Dataset, Ecommerce Image Dataset, Furniture Dataset

FAQ

Do your images include CLIP embeddings or just tags?

ImageHub provides CLIP-derived tag labels (attribute strings) — not raw embedding vectors. If you need pre-computed embeddings, you can run your CLIP model over the downloaded images using the provided metadata as filtering input.

How do I get images from the same category across multiple retailers?

Submit a custom bundle request specifying the category and select "all sources" — we'll pull matching images across all retailers in that category.

Is this suitable for training a "shop the look" model?

Yes — the multi-source, multi-angle structure is well-suited. For best results, request lifestyle images for the "look" side and primary product images for the "shop" side.

How many images per category are available?

Ranges from 16,000+ (furniture) to 84,000+ (toys & games) depending on category. See the full catalog for current counts.