Visual Search Dataset – Diverse Product Images for Similarity & Retrieval Models

Visual search is the fastest-growing discovery method in ecommerce. Shoppers photograph or upload an image and expect the system to surface identical or visually similar products within milliseconds. Building a reliable visual search pipeline requires a large, diverse, well-labeled training dataset — and that's exactly what the ImageHub visual search dataset provides.

What Visual Search Models Need

Effective visual search depends on training data that mirrors the variety of real-world query images. Shoppers photograph products from odd angles, in different lighting conditions, and sometimes with backgrounds. A model trained only on clean studio shots won't handle these conditions well.

• Varied angles — front, side, back, and lifestyle shots reflecting real camera positions.
• Lighting diversity — studio white background, natural light lifestyle, and catalog lighting mixed.
• Fine-grained attribute labels — color, pattern, material, and style for similarity-aware search.
• Negative examples — different products in the same category for contrastive training pairs.

How Teams Use This Dataset

Training Visual Embedding Models

Teams fine-tune contrastive models (CLIP, SigLIP, ResNet-based triplet networks) using the dataset. The existing tag structure provides both category-level labels for classification and attribute-level labels for retrieval quality metrics. The multi-angle coverage per SKU means you can form positive pairs trivially.

Benchmarking Retrieval Systems

The dataset is useful as a retrieval benchmark: each SKU has a known ground-truth set of relevant images. Teams measure recall@k and precision@k to compare embedding architectures without needing to collect evaluation data separately.

Re-ranking & Attribute Filtering

Post-retrieval re-rankers that use attribute signals (same color, same material) improve conversion by showing visually similar items that also match inferred shopper intent. Train the re-ranker on the attribute labels in the dataset to teach it which dimensions matter most per category.

Dataset Details

• CLIP semantic tags at the image level — pre-computed and ready to use as embedding supervision
• Category and sub-category labels for coarse and fine-grained retrieval
• Identity grouping: multiple images per SKU — essential for positive-pair generation
• Diverse sources across fashion, furniture, beauty, health, and electronics

Get Started

Download a free sample dataset from the datasets page to evaluate the structure and tag quality. Once you're satisfied, request a full category or multi-category bundle from the catalog.

Browse by Category

Explore our image dataset organized by product category: