srpone/look-bench

LookBench: A Live and Holistic Fashion Image Retrieval Benchmark

LookBench is a large-scale, open benchmark for fashion image retrieval, designed to evaluate modern vision and vision–language models under realistic, contamination-aware settings. The benchmark emphasizes live data, domain diversity, and holistic retrieval tasks spanning both single-item and outfit-level scenarios.

This dataset accompanies the paper LookBench: A Live and Holistic Open Benchmark for Fashion Image Retrieval.

project page

code

🎯 Motivation

Existing fashion retrieval benchmarks often suffer from:

• Significant test–training contamination
• Over-reliance on clean studio product images
• Limited support for outfit-level and real-world queries

LookBench addresses these limitations by introducing live, recently collected images, street-style outfit queries, and AI-generated images, enabling more realistic and forward-looking evaluation.

---

📦 Dataset Overview

LookBench consists of multiple subsets reflecting different image sources and retrieval difficulties.
Each subset is constructed as a query–corpus retrieval benchmark, where query images are matched against a large gallery.

Subsets (from Table 1 in the paper)

Subset Name	Image Source	Retrieval Type	Difficulty	#Queries	#Corpus
RealStudioFlat	Real studio flat-lay product images	Single-item	Easy	1,011	62,226
AIGen-Studio	AI-generated studio images	Single-item	Medium	192	59,254
RealStreetLook	Real street outfit images	Multi-item	Hard	1,000	61,553
AIGen-StreetLook	AI-generated street outfit images	Multi-item	Hard	160	58,846

---

🧠 Tasks

LookBench supports two primary retrieval tasks:

1. Single-Item Retrieval

Given a query image containing a single fashion item, retrieve the exact matching product from the corpus.

2. Multi-Item (Outfit) Retrieval

Given a street-style image containing multiple fashion items, retrieve all corresponding products from the corpus.

These tasks reflect real-world fashion search and recommendation scenarios.

---

🧾 Data Format

Each dataset subset contains:

• Query split: images used as retrieval queries
• Corpus split: candidate images used as the retrieval gallery

Each sample may include the following fields (subset-dependent):

• image: Input fashion image
• category: Fashion category label
• bbox: Bounding box of the fashion item
• item_id: Unique product identifier
• task: Retrieval task type
• difficulty: Difficulty level

---

🚀 How to Use

Load the Dataset

You can load LookBench using the 🤗 Datasets library:

from datasets import load_dataset

dataset = load_dataset("srpone/look-bench")
print(dataset)

Citation

@article{gao2026lookbench,
      title={LookBench: A Live and Holistic Open Benchmark for Fashion Image Retrieval}, 
      author={Chao Gao and Siqiao Xue and Yimin Peng and Jiwen Fu and Tingyi Gu and Shanshan Li and Fan Zhou},
      year={2026},
      url={https://arxiv.org/abs/2601.14706}, 
      journal= {arXiv preprint arXiv:2601.14706},
}