LookBench
A Live and Holistic Open Benchmark for Fashion Image Retrieval
Last updated: June 30, 2026
📄 [2026-06-30]: New submission from ZooClaw.ai: ZooClaw-FashionSigLIP2 — a SigLIP2-base model fine-tuned with knowledge distillation for robust fashion retrieval, evaluated on text-image retrieval tasks (ZooClaw-Fashion long/short and H&M). Paper · Model 🎉
🎯 [2026-06-04]: New submission from Kuaishou Team: Tianmu-MERE — a multimodal embedding model for product understanding and retrieval in e-commerce, achieving strong results across all subtasks! Model weights are now open-sourced on HuggingFace. 🏆
🚀 [2026-01-20]: Initial release of LookBench v2601 with 4 diverse subtasks across AI-generated and real-world scenarios, and a new open-source model GR-Lite! 🌟
🔥 [2026-01-20]: Release of GR-Pro (proprietary) and GR-Lite (open-source) models achieving state-of-the-art performance! 🎉
LookBench is a live, holistic, and challenging benchmark for fashion image retrieval in real e-commerce settings. Unlike static benchmarks that are vulnerable to data contamination, LookBench features continuously refreshing samples, diverse retrieval intents across multiple difficulty levels, and attribute-supervised evaluation with over 100 visually grounded properties.
The current release (v2601) comprises approximately 2,300 queries, each evaluated against a carefully curated retrieval corpus of about 60,000 images per task:
| Dataset | Image Source | # Retro Items | Difficulty | # Queries / Corpus |
|---|---|---|---|---|
| RealStudioFlat | Real studio flat-lay product photos | Single | Easy | 1,011 / 62,226 |
| AIGen-Studio | AI-generated lifestyle studio images | Single | Medium | 192 / 59,254 |
| RealStreetLook | Real street outfit photos | Multi | Hard | 1,000 / 61,553 |
| AIGen-StreetLook | AI-generated street outfit compositions | Multi | Hard | 160 / 58,846 |
Each evaluation set is assessed using Coarse Recall, Fine Recall, and nDCG at @1, @5, @10, and @20.
Results of different models. The best-performing model in each metric is in-bold, and the second best is underlined.
GR-Pro and GR-Lite are our proprietary and open-source models respectively.
ZooClaw-Fashion and H&M numbers are from Xue & Xu, 2026. ZooClaw-Fashion reports R@1 and R@10 (long & short query); H&M reports R@10 + MRR@10. The ZooClaw-Fashion evaluation dataset will be released shortly at srpone/zooclaw-fashion-eval.
@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},
}
@article{xue2026zooclaw,
title={ZooClaw-FashionSigLIP2: Distilled Fine-tuning for Robust Fashion Retrieval},
author={Siqiao Xue and Chunxue Xu},
year={2026},
url={https://arxiv.org/abs/2606.27708},
journal={arXiv preprint arXiv:2606.27708},
}