Hopit AI — fashion retrieval, measured properly

We build the MODA family of fashion retrieval models and benchmark them the way we wish everyone did: full corpus, one shared harness, every competitor under identical protocol — including the cells we lose.

Full benchmark page · Interactive demo · Code & methodology · hopit.ai · Book a call

Text-to-image — models under 250M parameters

The size class most deployments use. Full-corpus MAP@10; best per row bold 🥇, second 🥈:

Every benchmark in this class is led by a MODA-family model except one, where frozen FashionSigLIP keeps a 0.8% edge. MODA Pro Lite owns catalog and title search (KAGL +10.2%, Polyvore +7.3% over MODA, both significant) from one checkpoint, no serving recipe; MODA owns captions and instance retrieval (4 of 6 wins over FashionSigLIP significant under paired bootstrap).

Text-to-image — all systems, including larger models

Full-corpus MAP@10 through one shared pipeline, identical preprocessing per model, no gallery subsampling anywhere. Best per row in bold with 🥇, second 🥈.

MODA Pro: rank 1 or 2 on every row, and on 9 of 10 cells once the H&M and ZooClaw-Fashion registers are included — the only system with no bad benchmark (+6.9% mean over MODA on these six, peak +12.9%). MODA Pro Lite is the strongest open single model ≤250M on catalog search: KAGL +10.2% and Polyvore +7.3% over MODA (both significant), from one checkpoint with no serving machinery. Full page with ranks, ops costs and losses: benchmark page.

Image-to-image retrieval — LookBench Fine R@1

†same-harness reruns; our harness reproduces Tianmu's published 66.20 within 0.5pt. A 203M open model above a closed commercial system and a 1.24B model.

The MODA family

Why trust these numbers

• Full corpus only — no subsampled galleries; screening runs are never mixed with full-corpus rows.
• One harness — every model, ours and competitors', runs identical preprocessing and protocol.
• Losses shown — every model card links the cells it loses; paired-bootstrap CIs where significance is claimed.
• Reproducible — code, harness, and per-cell receipts in the public repo.

Which model should I use?

All models serve on CPU. Each card links the benchmarks it loses, too.

Quick start

Text to image (plain bi-encoder, works with any vector DB):

import open_clip

model, _, preprocess = open_clip.create_model_and_transforms("hf-hub:HopitAI/moda-pro-lite")
tokenizer = open_clip.get_tokenizer("hf-hub:HopitAI/moda-pro-lite")

Image to image, for finding visually similar products:

import open_clip

model, _, preprocess = open_clip.create_model_and_transforms("hf-hub:HopitAI/moda-fashion-distilled")

MODA's full multi-view retrieval recipe:

pip install "git+https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m"

from moda_fashionsiglip_multiview import ModaFashionSigLIP

retriever = ModaFashionSigLIP.from_pretrained()
index = retriever.build_index(image_paths, item_ids=item_ids)
results = retriever.search("red floral summer dress", index, top_k=5)[0]

Use cases

• Search a fashion catalog with a natural-language query.
• Find visually similar products in a fashion catalog.
• Match street-style looks to shoppable items.
• Deduplicate product images across marketplaces.
• Build embedding indexes for ecommerce search and recommendations.