Fashion embeddings (local)

General-purpose embedders are not tuned for catalogue photos and product titles. These fashion-specific encoders are, and they plug into the embed service next to EmbeddingGemma 2 and pplx-embed-v2-late. Typical uses: search a catalogue by description or by photo, find "more like this", deduplicate product shots, group lookbooks.

CLI model	Repo	Modes	Licence	Notes
marqo-fashionsiglip	Marqo/marqo-fashionSigLIP	text + image	Apache-2.0	ViT-B/16 SigLIP, ~0.2B, 768-d. Marqo reports up to +57% MRR / recall over FashionCLIP (vendor-reported). The most-downloaded fashion encoder.
moda-fashion-distilled	HopitAI/moda-fashion-distilled	image only	MIT	MODA's image-to-image checkpoint, distilled from a 3-model ensemble. 67.63 fine R@1 on LookBench vs 63.84 for FashionSigLIP (vendor-reported). The card documents no text usage, so text is blocked here.
moda-pro-lite	HopitAI/moda-pro-lite	text + image	Apache-2.0	213M text-to-image encoder for catalogue and title search. Bare encoder only — MODA's "Plus" multi-view recipe and ANN helpers are not wired.
zooclaw-fashionsiglip2	srpone/zooclaw-fashionsiglip2	text + image	Apache-2.0	SigLIP2-base fine-tune (~375M), 384 px images, 64 text tokens, arXiv 2606.27708. The card lists no benchmarks.

All vectors are L2-normalised, so cosine similarity is a dot product. Benchmark numbers above come from the model authors and were not reproduced here.

Requirements

pip install opentryon[local]
pip install open_clip_torch      # Marqo and MODA (ZooClaw uses plain transformers)

Pass --model-id to point at a local snapshot or a different repo.

CLI

# Catalogue search by text
opentryon embed --model marqo-fashionsiglip \
  --image boot1.jpg boot2.jpg boot3.jpg --query "black leather ankle boots"

# "More like this" by photo
opentryon embed --model moda-fashion-distilled \
  --image catalogue/*.jpg --query-image customer_photo.jpg

# Embed product titles and photos into one space
opentryon embed --model zooclaw-fashionsiglip2 \
  --text "linen wrap dress" "wool overcoat" --image dress.jpg

Vectors are saved to outputs/embed_<model>.npz (emb_0 … emb_{n-1}, text first, then images). With --query or --query-image the command also prints a cosine score per input. --query and --query-image are mutually exclusive.

Python

from tryon.models.fashion_embed import FashionOpenCLIPAdapter, ZooClawFashionSigLIP2Adapter

adapter = FashionOpenCLIPAdapter(variant="marqo-fashionsiglip")   # or moda-fashion-distilled / moda-pro-lite
out = adapter.embed(image=["a.jpg", "b.jpg"], query="red satin evening gown")
print(out["scores"], out["embeddings"].shape)

zoo = ZooClawFashionSigLIP2Adapter()
out = zoo.embed(text=["linen shirt"], image=["shirt.jpg"], query_image="query.jpg")

Notes

• MCP tools: embed_marqo_fashionsiglip, embed_moda_fashion_distilled, embed_moda_pro_lite, embed_zooclaw_fashionsiglip2.
• Planner pins: marqo fashionsiglip, moda pro lite, moda fashion, zooclaw. embed stays named-model-only.
• No GPU was available when these were integrated; the adapters are tested against stubbed open_clip / transformers models, not real weights. Run a quick sanity retrieval on your own catalogue before relying on any of them.
• Not integrated: the MODA multi-view "Plus" recipe, moda-fashion-matryoshka / -vision-fp16 variants, and the Trendyol e-commerce DINOv2 embedder (CC BY-SA). See Integrate next.