HopitAI/moda-fashionsiglip-multiview-203m

MODA

The strongest open fashion retrieval system at 203M — a zero-new-parameter serving recipe over frozen FashionSigLIP that wins 4 of 6 full-corpus benchmarks against its own base model, with paired-bootstrap significance.

By Hopit AI. Repo name kept as moda-fashionsiglip-multiview-203m for link stability; the product name is MODA.

A downloadable implementation of the zero-additional-parameter retrieval architecture that produced four statistically significant full-corpus text-to-image wins over official FashionSigLIP.

This is a retrieval system, not a newly trained checkpoint. It downloads the unchanged Apache-2.0 `Marqo/marqo-fashionSigLIP` weights and applies a frozen query, gallery, and scoring recipe.

Book a call to evaluate on your own catalog · hopit.ai

The MODA family

All numbers full-corpus through one shared harness, measured alongside FashionSigLIP, SigLIP-SO400M (878M) and ZooClaw under identical protocol — including the cells each model loses. Full table: benchmark page.

Result

The same global recipe was evaluated on all six datasets using full-corpus MAP@10 and 10,000 paired-bootstrap samples.

Delta is relative to the FashionSigLIP baseline. The confidence interval is on the absolute MAP@10 difference, which is what the bootstrap resamples.

Accurate summary: 4/6 significant wins, positive point estimates on 6/6, and zero significant losses.

Install

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

The first model load downloads the upstream FashionSigLIP checkpoint.

Quick start

from pathlib import Path

from moda_fashionsiglip_multiview import ModaFashionSigLIP

images = sorted(Path("my_catalog").glob("*.jpg"))

retriever = ModaFashionSigLIP.from_pretrained()  # CUDA → MPS → CPU
index = retriever.build_index(
    images,
    item_ids=[path.name for path in images],
)

results = retriever.search(
    "red floral summer dress",
    index,
    top_k=5,
)[0]

for result in results:
    print(result.rank, result.score, result.item_id)

Apple Silicon is supported explicitly:

retriever = ModaFashionSigLIP.from_pretrained(device="mps")

The installation provides a complete folder-search command:

moda-fashion-search \
  --gallery ./my_catalog \
  --query "black leather ankle boots" \
  --top-k 10 \
  --device mps

Standalone inference script

inference.py runs from a plain clone of this repository without installing the package. It covers the three things people usually want first: ranking a folder against a query, embedding images, and embedding a query.

# rank a folder of product images against a text query
python inference.py --gallery ./my_catalog --query "red floral summer dress"

# embed images (768-d parent vectors)
python inference.py --image img1.jpg img2.jpg

# embed a query
python inference.py --query "black leather ankle boots"

# build once, reuse the saved index later
python inference.py --gallery ./my_catalog --query "navy linen shirt" \
  --save-index ./catalog_index
python inference.py --load-index ./catalog_index --query "wool coat"

Save and reload a gallery

The three gallery routes are stored with SafeTensors; no pickle loading is required.

index.save_pretrained("./catalog_index")

from moda_fashionsiglip_multiview import GalleryIndex

index = GalleryIndex.from_pretrained("./catalog_index")
results = retriever.search("navy linen shirt", index)[0]

Architecture

Query

q_raw    = FashionSigLIP(query)
q_prompt = FashionSigLIP("a fashion product photo of " + query)
q        = normalize(q_raw + 0.25 × q_prompt)

Gallery

d_official = FashionSigLIP(original image)
d_pad      = FashionSigLIP(gray-128 square-padded image)
d_crop     = FashionSigLIP(center-square-cropped image)

d_parent = normalize(
    d_official + 0.25 × d_pad + 0.25 × d_crop
)

The index stores d_parent, d_pad, and d_crop.

Score

s_parent = cosine(q, d_parent)
s_pad    = cosine(q, d_pad)
s_crop   = cosine(q, d_crop)

score = 0.9 × s_parent
      + 0.1 × max(s_parent, s_pad, s_crop)

Deployment profile

ModaFashionSigLIP.search() performs exact, chunked scoring so the example is easy to reproduce. At production scale, search three ANN indexes, union their candidates, and apply the identical fusion formula to those candidates. GalleryIndex.save_pretrained() preserves FP32 vectors by default for the most conservative reproduction; production indexes may store FP16 vectors to reach the storage figure above.

Evaluation disclosure

• —Task: text-to-image retrieval.
• —Primary metric: full-corpus MAP@10.
• —Confidence intervals: paired bootstrap with 10,000 samples.
• —One global recipe was used across all six datasets.
• —The exact 0.10 late-fusion blend was selected on external OpenVTON and leakage-audited GLAMI development data, not on target examples, images, or qrels.
• —Previous target aggregate results were already known by benchmark iteration six.
• —This is not a fresh independent blind SOTA evaluation.
• —Atlas and DeepFashion Multimodal are numerical improvements, not significant wins.

The frozen result summary and exact public recipe are included in this repository under benchmark_results/ and config.json.

For a cloned repository, python example_retrieval.py ... invokes the same installed command-line implementation.

Intended use

• —Fashion catalog text-to-image search.
• —Reproducing the deterministic architecture on a private image collection.
• —Studying multi-view preprocessing and conservative score fusion.

Limitations

• —Three image encodings are needed when building a gallery.
• —The index uses three vectors and three retrieval routes per product.
• —The query path performs two text encodings.
• —Results can differ if the upstream checkpoint, OpenCLIP preprocessing, image decoding, corpus construction, or metric implementation changes.
• —Fashion retrieval benchmarks do not establish performance for people recognition, biometric use, or unrelated visual domains.

Licenses

The wrapper and retrieval code in this repository are MIT licensed. The referenced Marqo FashionSigLIP checkpoint is distributed by Marqo under Apache-2.0. No upstream weights are duplicated here.