# MODA: open fashion retrieval benchmark and models

> An open benchmark and model family for fashion retrieval by Hopit AI. Every
> number is measured at full corpus through one shared harness, with competitor
> models run under identical protocol and every losing cell disclosed.

MODA covers both halves of fashion search: finding products from a written
description (text-to-image) and from a photo (image-to-image). Code, harness
and result artifacts are public and MIT licensed.

## The model family

- MODA (203M params, 768-d). Open source + open weights. A zero-new-parameter
  serving recipe over the frozen Apache-2.0 Marqo/marqo-fashionSigLIP
  checkpoint: three image views per product, two prompt-blended query
  encodings, late fusion. Four of six full-corpus wins over its own base model
  are statistically significant under a paired bootstrap. Leads the sub-250M
  class on caption-style and exact-item retrieval.
  https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m
- MODA Pro Lite (213M params, 768-d). Open weights, Apache-2.0. A trained
  fashion encoder on a SigLIP2-base-384 backbone with a verified
  fashion-vocabulary build (44,646 pieces, bit-identical outputs to the
  untrimmed model on 18,000 test texts). Beats MODA at full corpus on KAGL
  (+10.2%) and Polyvore (+7.3%), both significant under a paired bootstrap
  (10,000 resamples), as a plain bi-encoder: one
  vector per image, any vector database, no serving recipe.
  https://huggingface.co/HopitAI/moda-pro-lite
- MODA Pro. Closed, hosted via the Hopit API. Rank 1 or 2 on 9 of 10
  text-to-image benchmark cells across three venues; +6.9% mean over MODA on
  the six academic sets at full corpus (peak +12.9%). The only system in the
  comparison without a bad benchmark, at single-model query latency.
- MODA-SigLIP-Distilled (203M, 768-d). Open weights. Image-to-image retrieval.
  LookBench Fine Recall@1 67.63, the top open model, above a 1.24B-parameter
  model. https://huggingface.co/HopitAI/moda-fashion-distilled
- Variants: Matryoshka (64–768 dimensions selectable at query time; 256-d
  matches 768-d on LookBench Fine R@1), 512d, and an FP16 vision tower
  (186 MB) for edge deployment.

## Text-to-image results (full corpus, MAP@10)

Six academic benchmarks, no gallery subsampling. Systems compared: Marqo
FashionSigLIP (203M), MODA (203M), SigLIP-SO400M-384 (878M),
ZooClaw-FashionSigLIP2 (375M), MODA Pro Lite (213M), MODA Pro.

| Benchmark (corpus) | FashionSigLIP | MODA | SO400M | ZooClaw | Pro Lite | MODA Pro |
|---|---|---|---|---|---|---|
| KAGL (44K) | 0.2769 | 0.2890 | 0.3370 | 0.2951 | 0.3185 | 0.3263 |
| Polyvore (94K) | 0.3665 | 0.3726 | 0.4378 | 0.3804 | 0.3997 | 0.4088 |
| Atlas (78K) | 0.1826 | 0.1884 | 0.2309 | 0.1583 | 0.1945 | 0.2053 |
| Fashion200K (202K) | 0.1858 | 0.1947 | 0.1353 | 0.1775 | 0.1802 | 0.2101 |
| DeepFashion In-Shop (53K) | 0.1587 | 0.1703 | 0.1695 | 0.1024 | 0.1031 | 0.1762 |
| DeepFashion Multimodal (43K) | 0.0148 | 0.0147 | 0.0079 | 0.0099 | 0.0118 | 0.0144 |

MODA Pro wins Fashion200K and DeepFashion In-Shop outright and is second on
KAGL, Polyvore and Atlas, where the 878M SigLIP-SO400M leads. FashionSigLIP
holds DeepFashion Multimodal by 0.8%.

## Image-to-image results (LookBench, Fine Recall@1)

| Model | Params | Fine R@1 |
|---|---|---|
| MODA-SigLIP-Distilled | 203M | 67.63 |
| GR-Pro (closed) | n/a | 67.38 |
| Tianmu-MERE | 1.24B | 65.99 |
| GR-Lite | n/a | 65.71 |
| FashionSigLIP (our reproduction) | 203M | 63.84 |

Tianmu-MERE was re-run through this harness and reproduces its own published
subset scores within 0.5 points, which validates the harness.

## Method and limitations

- Full corpus only. Subsampled screening runs are never mixed with full-corpus
  rows in the same comparison.
- One harness, identical preprocessing per model. Published baselines are
  reproduced before any comparison (FashionSigLIP within +0.08%).
- Losses are published alongside wins; each model card links the benchmarks it
  loses.
- The evaluation is self-declared and non-blind: aggregate results were known
  during development. It should not be described as an independently verified
  state-of-the-art result.
- Absolute values on purchase-grounded sets are low by construction (one bought
  item per query against a large catalog); relative differences carry the
  signal.

## Links

- Benchmark page: https://hopit-ai.github.io/Moda/
- Code and harness: https://github.com/hopit-ai/Moda
- Reproduction recipe (exact commands, expected numbers, known variance):
  https://github.com/hopit-ai/Moda/blob/main/REPRODUCE.md
- Book a call: https://calendly.com/arkid_/new-meeting?back=1
- Models: https://huggingface.co/HopitAI
- Company: https://hopit.ai
- Research notes: https://hopitai.substack.com/