ArkidMitra commited on
Commit
6ed089f
Β·
verified Β·
1 Parent(s): b736a5b

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +119 -172
README.md CHANGED
@@ -1,185 +1,132 @@
1
- ---
2
- title: README
3
- emoji: πŸ§₯
4
- colorFrom: indigo
5
- colorTo: green
6
- sdk: static
7
- pinned: false
8
- license: mit
9
- short_description: SOTA fashion retrieval, measured and open.
10
- ---
11
-
12
- <section>
13
- <h1>Hopit AI</h1>
14
- <p><b>Fashion AI infrastructure.</b> We build the MODA family of fashion retrieval
15
- models and benchmark them the way we wish everyone did: <b>full corpus, one shared
16
- harness, every competitor under identical protocol, including the cells we lose.</b></p>
17
- <p>
18
- <a href="https://hopit.ai">hopit.ai</a> Β·
19
- <a href="https://hopit-ai.github.io/Moda/">Benchmarks</a> Β·
20
- <a href="https://github.com/hopit-ai/Moda">Code</a> Β·
21
- <a href="https://huggingface.co/spaces/HopitAI/moda-fashion-search">Demo</a> Β·
22
- <a href="https://calendly.com/arkid_/new-meeting?back=1">Book a call</a>
23
- </p>
24
- <p><b>Start here.</b> Text search: <a href="https://huggingface.co/HopitAI/moda-pro-lite">MODA Pro Lite</a> (213M, open weights).
25
- Image search: <a href="https://huggingface.co/HopitAI/moda-fashion-distilled">MODA-SigLIP-Distilled</a> (203M, #1 open on LookBench).
26
- Both are plain bi-encoders that work with any vector database.</p>
27
- </section>
28
-
29
- <section>
30
- <h2>Text-to-image β€” models under 250M parameters</h2>
31
- <p>The size class most deployments use. Full-corpus MAP@10; best per row bold πŸ₯‡, second πŸ₯ˆ:</p>
32
- <table>
33
- <tr><th>benchmark (corpus)</th><th><a href="https://huggingface.co/Marqo/marqo-fashionSigLIP">FashionSigLIP</a> 203M</th><th>MODA 203M</th><th>MODA Pro Lite 213M</th></tr>
34
- <tr><td>KAGL (44K)</td><td>0.2769</td><td>0.2890 πŸ₯ˆ</td><td><b>0.3185</b> πŸ₯‡</td></tr>
35
- <tr><td>Polyvore (94K)</td><td>0.3665</td><td>0.3726 πŸ₯ˆ</td><td><b>0.3997</b> πŸ₯‡</td></tr>
36
- <tr><td>Atlas (78K)</td><td>0.1826</td><td>0.1884 πŸ₯ˆ</td><td><b>0.1945</b> πŸ₯‡</td></tr>
37
- <tr><td>Fashion200K (202K)</td><td>0.1858 πŸ₯ˆ</td><td><b>0.1947</b> πŸ₯‡</td><td>0.1802</td></tr>
38
- <tr><td>DeepFashion In-Shop (53K)</td><td>0.1587 πŸ₯ˆ</td><td><b>0.1703</b> πŸ₯‡</td><td>0.1031</td></tr>
39
- <tr><td>DeepFashion Multimodal (43K)</td><td><b>0.0148</b> πŸ₯‡</td><td>0.0147 πŸ₯ˆ</td><td>0.0118</td></tr>
40
- </table>
41
- <p>
42
- Every benchmark in this class is led by a MODA-family model except one, where
43
- frozen FashionSigLIP keeps a 0.8% edge.
44
- <b><a href="https://huggingface.co/HopitAI/moda-pro-lite">MODA Pro Lite</a> owns catalog
45
- and title search</b> (KAGL +10.2%, Polyvore +7.3% over MODA, both significant) from one
46
- checkpoint, no serving recipe; <b>MODA owns captions and instance retrieval</b>
47
- (4 of 6 wins over FashionSigLIP significant under paired bootstrap).
48
- </p>
49
- </section>
50
-
51
- <section>
52
- <h2>Text-to-image β€” all systems, including larger models</h2>
53
- <p>
54
- Full-corpus MAP@10 through one shared pipeline, identical preprocessing per
55
- model, no gallery subsampling anywhere. Best per row in bold with πŸ₯‡, second πŸ₯ˆ.
56
- </p>
57
- <table>
58
- <tr><th>benchmark (corpus)</th><th><a href="https://huggingface.co/Marqo/marqo-fashionSigLIP">FashionSigLIP</a><br>203M</th><th>MODA<br>203M</th><th><a href="https://huggingface.co/timm/ViT-SO400M-14-SigLIP-384">SO400M</a><br>878M</th><th><a href="https://huggingface.co/srpone/zooclaw-fashionsiglip2">ZooClaw</a><br>375M</th><th>MODA Pro Lite<br>213M</th><th>MODA Pro<br>hosted</th></tr>
59
- <tr><td>KAGL (44K)</td><td>0.2769</td><td>0.2890</td><td><b>0.3370</b> πŸ₯‡</td><td>0.2951</td><td>0.3185</td><td>0.3263 πŸ₯ˆ</td></tr>
60
- <tr><td>Polyvore (94K)</td><td>0.3665</td><td>0.3726</td><td><b>0.4378</b> πŸ₯‡</td><td>0.3804</td><td>0.3997</td><td>0.4088 πŸ₯ˆ</td></tr>
61
- <tr><td>Atlas (78K)</td><td>0.1826</td><td>0.1884</td><td><b>0.2309</b> πŸ₯‡</td><td>0.1583</td><td>0.1945</td><td>0.2053 πŸ₯ˆ</td></tr>
62
- <tr><td>Fashion200K (202K)</td><td>0.1858</td><td>0.1947 πŸ₯ˆ</td><td>0.1353</td><td>0.1775</td><td>0.1802</td><td><b>0.2101</b> πŸ₯‡</td></tr>
63
- <tr><td>DeepFashion In-Shop (53K)</td><td>0.1587</td><td>0.1703 πŸ₯ˆ</td><td>0.1695</td><td>0.1024</td><td>0.1031</td><td><b>0.1762</b> πŸ₯‡</td></tr>
64
- <tr><td>DeepFashion Multimodal (43K)</td><td><b>0.0148</b> πŸ₯‡</td><td>0.0147 πŸ₯ˆ</td><td>0.0079</td><td>0.0099</td><td>0.0118</td><td>0.0144</td></tr>
65
- </table>
66
- <p>
67
- <b>MODA Pro: rank 1 or 2 on every row, and on 9 of 10 cells</b> once the
68
- H&amp;M and ZooClaw-Fashion registers are included β€” the only system with no
69
- bad benchmark (+6.9% mean over MODA on these six, peak +12.9%).
70
- <b>MODA Pro Lite</b> is the strongest open single model ≀250M on catalog
71
- search: KAGL <b>+10.2%</b> and Polyvore <b>+7.3%</b> over MODA (both significant), from one checkpoint with no serving machinery. Full page with ranks,
72
- ops costs and losses: <a href="https://hopit-ai.github.io/Moda/">benchmark page</a>.
73
- </p>
74
- </section>
75
-
76
- <section>
77
- <h2>Image-to-image retrieval β€” LookBench Fine R@1</h2>
78
- <table>
79
- <tr><th>model</th><th>params</th><th>Fine R@1</th></tr>
80
- <tr><td><b><a href="https://huggingface.co/HopitAI/moda-fashion-distilled">MODA-SigLIP-Distilled</a></b></td><td>203M</td><td><b>67.63</b> πŸ₯‡</td></tr>
81
- <tr><td><a href="https://serendipityoneinc.github.io/look-bench-page/">GR-Pro</a> (closed)</td><td>n/a</td><td>67.38</td></tr>
82
- <tr><td><a href="https://huggingface.co/TianmuLab/Tianmu-MERE">Tianmu-MERE</a></td><td>1.24B</td><td>65.99†</td></tr>
83
- <tr><td><a href="https://huggingface.co/Marqo/marqo-fashionSigLIP">FashionSigLIP</a></td><td>203M</td><td>63.84†</td></tr>
84
- </table>
85
- <p>
86
- †same-harness reruns; our harness reproduces Tianmu's published 66.20 within
87
- 0.5pt. A 203M open model above a closed commercial system and a 1.24B model.
88
- </p>
89
- </section>
90
-
91
- <section>
92
- <h2>The MODA family</h2>
93
- <table>
94
- <tr><th>model</th><th>what it is</th><th>availability</th></tr>
95
- <tr>
96
- <td><a href="https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m"><b>MODA</b></a> (203M)</td>
97
- <td>zero-new-parameter serving recipe over frozen FashionSigLIP; 4/6 statistically significant full-corpus wins over its own base model</td>
98
- <td>open source + open weights</td>
99
- </tr>
100
- <tr>
101
- <td><a href="https://huggingface.co/HopitAI/moda-pro-lite"><b>MODA Pro Lite</b></a> (213M)</td>
102
- <td>trained encoder (verified fashion-vocab build); beats MODA on catalog search as a plain bi-encoder β€” no recipe required</td>
103
- <td>open weights</td>
104
- </tr>
105
- <tr>
106
- <td><b>MODA Pro</b></td>
107
- <td>our hosted retrieval system β€” rank 1 or 2 on 9 of 10 benchmark cells at single-model query latency and cost</td>
108
- <td>closed Β· <a href="https://hopit.ai">hosted</a></td>
109
- </tr>
110
- <tr>
111
- <td><a href="https://huggingface.co/HopitAI/moda-fashion-distilled"><b>MODA-SigLIP-Distilled</b></a> (203M)</td>
112
- <td>image-to-image specialist β€” #1 open model on LookBench</td>
113
- <td>open weights (+ <a href="https://huggingface.co/HopitAI/moda-fashion-matryoshka">matryoshka</a>, <a href="https://huggingface.co/HopitAI/moda-fashion-distilled-512d">512d</a>, <a href="https://huggingface.co/HopitAI/moda-fashion-vision-fp16">fp16-vision</a>, <a href="https://huggingface.co/HopitAI/moda-fashion-crossdomain">cross-domain</a> variants)</td>
114
- </tr>
115
- </table>
116
- </section>
117
-
118
- <section>
119
- <h2>Why trust these numbers</h2>
120
- <ul>
121
- <li><b>Full corpus only</b> β€” no subsampled galleries; screening runs are never mixed with full-corpus rows.</li>
122
- <li><b>One harness</b> β€” every model, ours and competitors', runs identical preprocessing and protocol.</li>
123
- <li><b>Losses shown</b> β€” every model card links the cells it loses; paired-bootstrap CIs where significance is claimed.</li>
124
- <li><b>Reproducible</b> β€” code, harness, and per-cell receipts in the <a href="https://github.com/hopit-ai/Moda">public repo</a>.</li>
125
- </ul>
126
- </section>
127
-
128
- <section>
129
- <h2>Which model should I use?</h2>
130
- <table>
131
- <tr><th>your workload</th><th>use</th><th>why</th></tr>
132
- <tr><td>catalog / title search</td><td><a href="https://huggingface.co/HopitAI/moda-pro-lite">moda-pro-lite</a></td><td>strongest ≀250M on catalog benchmarks; plain bi-encoder, any vector DB</td></tr>
133
- <tr><td>caption-style / exact-item search</td><td><a href="https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m">MODA</a></td><td>leads the class on Fashion200K, In-Shop, Multimodal</td></tr>
134
- <tr><td>best text search, zero integration</td><td><a href="https://hopit.ai">MODA Pro</a> (hosted)</td><td>rank 1–2 on 9 of 10 benchmarks, no bad benchmark</td></tr>
135
- <tr><td>visually similar products (image)</td><td><a href="https://huggingface.co/HopitAI/moda-fashion-distilled">moda-fashion-distilled</a></td><td>#1 open model on LookBench</td></tr>
136
- <tr><td>small index / edge</td><td><a href="https://huggingface.co/HopitAI/moda-fashion-matryoshka">matryoshka</a> @256d Β· <a href="https://huggingface.co/HopitAI/moda-fashion-vision-fp16">fp16</a></td><td>3Γ— smaller index at no loss Β· 186 MB vision tower</td></tr>
137
- </table>
138
- <p>All models serve on CPU. Each card links the benchmarks it loses, too.</p>
139
- </section>
140
-
141
- <section>
142
- <h2>Quick start</h2>
143
-
144
- Text to image (plain bi-encoder, works with any vector DB):
145
 
146
  ```python
147
- import open_clip
148
 
149
  model, _, preprocess = open_clip.create_model_and_transforms("hf-hub:HopitAI/moda-pro-lite")
150
  tokenizer = open_clip.get_tokenizer("hf-hub:HopitAI/moda-pro-lite")
 
 
 
 
 
 
151
  ```
152
 
153
- Image to image, for finding visually similar products:
 
154
 
155
- ```python
156
- import open_clip
157
 
158
- model, _, preprocess = open_clip.create_model_and_transforms("hf-hub:HopitAI/moda-fashion-distilled")
159
- ```
 
160
 
161
- MODA's full multi-view retrieval recipe:
162
 
163
- ```bash
164
- pip install "git+https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m"
165
- ```
 
 
 
 
 
 
 
166
 
167
- ```python
168
- from moda_fashionsiglip_multiview import ModaFashionSigLIP
169
 
170
- retriever = ModaFashionSigLIP.from_pretrained()
171
- index = retriever.build_index(image_paths, item_ids=item_ids)
172
- results = retriever.search("red floral summer dress", index, top_k=5)[0]
173
- ```
174
- </section>
175
-
176
- <section>
177
- <h2>Use cases</h2>
178
- <ul>
179
- <li>Search a fashion catalog with a natural-language query.</li>
180
- <li>Find visually similar products in a fashion catalog.</li>
181
- <li>Match street-style looks to shoppable items.</li>
182
- <li>Deduplicate product images across marketplaces.</li>
183
- <li>Build embedding indexes for ecommerce search and recommendations.</li>
184
- </ul>
185
- </section>
 
1
+ # Hopit AI
2
+
3
+ Fashion AI infrastructure β€” retrieval, trend intelligence and catalogue models for fashion
4
+ commerce, measured in public against open benchmarks.
5
+
6
+ ## The MODA family
7
+
8
+ | model | what it is | availability |
9
+ |---|---|---|
10
+ | **[MODA Pro Lite+](https://huggingface.co/HopitAI/moda-pro-lite)** (213M) | Trained fashion encoder with a calibrated serving recipe. Leads the ≀250M class on catalogue and title search. | open weights |
11
+ | **[MODA Duo](https://huggingface.co/HopitAI/moda-duo)** | Routes each query to the constituent that suits its shape. Beats either alone on a mixed workload, at one encoder per search. | open recipe Β· open weights |
12
+ | **[MODA](https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m)** (203M) | Zero-new-parameter serving recipe over frozen FashionSigLIP. Strongest of the open models on long descriptions. | open source Β· open weights |
13
+ | **MODA Pro** | Hosted retrieval system. Rank 1 or 2 on five of six full-corpus benchmarks. | closed Β· [hopit.ai](https://hopit.ai) |
14
+ | **[MODA-SigLIP-Distilled](https://huggingface.co/HopitAI/moda-fashion-distilled)** (203M) | Image-to-image specialist. #1 open model on LookBench. | open weights |
15
+
16
+ ## Text-to-image retrieval
17
+
18
+ MAP@10, full corpus, all ground-truth queries, one evaluator (`pytrec_eval map_cut.10`).
19
+ Every system in each size class is shown β€” ours and everyone else's, wins and losses.
20
+ Columns ordered by number of first places, then by top-two finishes.
21
+
22
+ ### Models at 250M parameters or fewer
23
+
24
+ | benchmark | [MODA Pro Lite+](https://huggingface.co/HopitAI/moda-pro-lite)<br>213M | [MODA](https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m)<br>203M | [FashionSigLIP](https://huggingface.co/Marqo/marqo-fashionSigLIP)<br>203M | [MODA Pro Lite](https://huggingface.co/HopitAI/moda-pro-lite)<br>213M |
25
+ |---|---:|---:|---:|---:|
26
+ | KAGL | **0.3201** | 0.2887 | 0.2769 | 0.3055 |
27
+ | Polyvore | **0.4049** | 0.3726 | 0.3664 | 0.3952 |
28
+ | Atlas | **0.1904** | 0.1862 | 0.1826 | 0.1814 |
29
+ | Fashion200K | 0.1846 | **0.1946** | 0.1858 | 0.1758 |
30
+ | DeepFashion In-Shop | 0.1026 | **0.1642** | 0.1586 | 0.0930 |
31
+ | DeepFashion Multimodal | 0.0133 | 0.0147 | **0.0148** | 0.0118 |
32
+
33
+ ### All systems
34
+
35
+ | benchmark | [SO400M](https://huggingface.co/timm/ViT-SO400M-14-SigLIP-384)<br>878M | MODA Pro<br>hosted | [MODA Duo](https://huggingface.co/HopitAI/moda-duo)<br>2 indexes | [MODA](https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m)<br>203M | [FashionSigLIP](https://huggingface.co/Marqo/marqo-fashionSigLIP)<br>203M | [MODA Pro Lite+](https://huggingface.co/HopitAI/moda-pro-lite)<br>213M | [MODA Pro Lite](https://huggingface.co/HopitAI/moda-pro-lite)<br>213M | [ZooClaw](https://huggingface.co/srpone/zooclaw-fashionsiglip2)<br>375M |
36
+ |---|---:|---:|---:|---:|---:|---:|---:|---:|
37
+ | KAGL | **0.3370** | 0.3221 | 0.3201 | 0.2887 | 0.2769 | 0.3201 | 0.3055 | 0.2951 |
38
+ | Polyvore | **0.4378** | 0.4106 | 0.4049 | 0.3726 | 0.3664 | 0.4049 | 0.3952 | 0.3804 |
39
+ | Atlas | **0.2309** | 0.2009 | 0.1904 | 0.1862 | 0.1826 | 0.1904 | 0.1814 | 0.1583 |
40
+ | Fashion200K | 0.1353 | **0.2065** | 0.1866 | 0.1946 | 0.1858 | 0.1846 | 0.1758 | 0.1775 |
41
+ | DeepFashion In-Shop | **0.1694** | 0.1663 | 0.1640 | 0.1642 | 0.1586 | 0.1026 | 0.0930 | 0.1024 |
42
+ | DeepFashion Multimodal | 0.0079 | 0.0137 | **0.0159** | 0.0147 | 0.0148 | 0.0133 | 0.0118 | 0.0099 |
43
+
44
+ ## Image-to-image retrieval
45
+
46
+ Finding visually similar products from a photo, on LookBench (2,345 queries,
47
+ [Gao et al., 2026](https://arxiv.org/abs/2601.14706)). Every system measured through one
48
+ harness; FashionSigLIP was reproduced here before any comparison was made.
49
+
50
+ | model | params | dim | Fine R@1 | nDCG@5 |
51
+ |---|---:|---:|---:|---:|
52
+ | [moda-fashion-distilled](https://huggingface.co/HopitAI/moda-fashion-distilled) | 203M | 768 | **67.63** | 53.85 |
53
+ | [moda-fashion-matryoshka](https://huggingface.co/HopitAI/moda-fashion-matryoshka) @256d | 203M | 256 | 67.42 | **57.48** |
54
+ | GR-Pro (SerendipityOne, closed) | n/a | 1024 | 67.38 | β€” |
55
+ | [moda-fashion-crossdomain](https://huggingface.co/HopitAI/moda-fashion-crossdomain) | 203M | 768 | 66.52 | 52.46 |
56
+ | [Tianmu-MERE](https://huggingface.co/TianmuLab/Tianmu-MERE) (Kuaishou) | 1.24B | 512 | 65.99 | 52.37 |
57
+ | [GR-Lite](https://huggingface.co/srpone/gr-lite) (SerendipityOne) | 303M | 1024 | 65.71 | β€” |
58
+ | [FashionSigLIP](https://huggingface.co/Marqo/marqo-fashionSigLIP) (our reproduction) | 203M | 768 | 63.84 | 49.63 |
59
+ | [FashionSigLIP](https://huggingface.co/Marqo/marqo-fashionSigLIP) (published) | 203M | 768 | 62.77 | 49.44 |
60
+
61
+ **Tianmu-MERE**, the top open model on the LookBench multi-attribute leaderboard, was
62
+ evaluated through this same harness and reproduces its own published subset scores within
63
+ 0.5 points β€” which independently validates the harness. moda-fashion-distilled leads it on
64
+ all four subsets, on Fine R@1 and nDCG@5 alike, at 6Γ— fewer parameters.
65
+
66
+ The configuration above was chosen from eight candidates ranked on LookBench itself, so the
67
+ headline carries the upward bias of a maximum over eight. It was independently reproduced
68
+ in-harness at 67.59, and the runner-up configurations land within 0.3 points.
69
+
70
+ ## Which model should I use?
71
+
72
+ | your workload | use | why |
73
+ |---|---|---|
74
+ | Catalogue / title search β€” `red floral maxi dress` | [moda-pro-lite](https://huggingface.co/HopitAI/moda-pro-lite) | leads the ≀250M class on KAGL, Polyvore and Atlas |
75
+ | Long descriptions, exact-item search | [moda-duo](https://huggingface.co/HopitAI/moda-duo) | routes long queries to the description-strong constituent |
76
+ | A mixture of both | [moda-duo](https://huggingface.co/HopitAI/moda-duo) | beats either constituent alone on a mixed workload |
77
+ | Agentic and multi-step search | MODA Pro ([hosted](https://hopit.ai)) | one API, no indexes to build or operate |
78
+ | Visually similar products from a photo | [moda-fashion-distilled](https://huggingface.co/HopitAI/moda-fashion-distilled) | #1 open model on LookBench |
79
+ | Same, but index size matters | [moda-fashion-matryoshka](https://huggingface.co/HopitAI/moda-fashion-matryoshka) | truncatable embeddings, 64–768d |
80
+ | Edge / mobile deployment | [moda-fashion-vision-fp16](https://huggingface.co/HopitAI/moda-fashion-vision-fp16) | 186 MB vision tower |
81
+
82
+ ## Quick start
83
+
84
+ ```bash
85
+ pip install open_clip_torch pillow
86
+ ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
87
 
88
  ```python
89
+ import open_clip, torch
90
 
91
  model, _, preprocess = open_clip.create_model_and_transforms("hf-hub:HopitAI/moda-pro-lite")
92
  tokenizer = open_clip.get_tokenizer("hf-hub:HopitAI/moda-pro-lite")
93
+ model.eval()
94
+
95
+ with torch.no_grad():
96
+ img = torch.nn.functional.normalize(model.encode_image(preprocess(image).unsqueeze(0)), dim=-1)
97
+ txt = torch.nn.functional.normalize(model.encode_text(tokenizer(["red floral maxi dress"])), dim=-1)
98
+ score = (txt @ img.T).item()
99
  ```
100
 
101
+ Each model repository ships a runnable script β€” `serving_ann.py` for the retrieval systems
102
+ (including the ANN index), `inference.py` for the image-to-image checkpoints.
103
 
104
+ ## Collections
 
105
 
106
+ - [Text-to-image retrieval](https://huggingface.co/collections/HopitAI/text-to-image-retrieval-6a7ad68d07fdebe00543a707)
107
+ - [Image-to-image retrieval](https://huggingface.co/collections/HopitAI/image-to-image-retrieval-6a7ad6aa12ecf6f5daec40f8)
108
+ - [Compact and edge deployment](https://huggingface.co/collections/HopitAI/compact-and-edge-deployment-6a7ad6ab5f5d8ac50882038d)
109
 
110
+ ## Models
111
 
112
+ | model | task | params | best for |
113
+ |---|---|---|---|
114
+ | [moda-duo](https://huggingface.co/HopitAI/moda-duo) | text β†’ image | recipe over two encoders | mixed workloads; routes per query, one encoder per search |
115
+ | [moda-pro-lite](https://huggingface.co/HopitAI/moda-pro-lite) | text β†’ image | 213M | catalogue and title search; ships its serving recipe |
116
+ | [moda-fashionsiglip-multiview-203m](https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m) | text β†’ image | 203M | long descriptions and exact-item retrieval |
117
+ | [moda-fashion-distilled](https://huggingface.co/HopitAI/moda-fashion-distilled) | image β†’ image | 203M | best overall visual similarity |
118
+ | [moda-fashion-matryoshka](https://huggingface.co/HopitAI/moda-fashion-matryoshka) | image β†’ image | 203M | truncatable embeddings, 64–768d |
119
+ | [moda-fashion-distilled-512d](https://huggingface.co/HopitAI/moda-fashion-distilled-512d) | image β†’ image | 203M | compact indexes |
120
+ | [moda-fashion-crossdomain](https://huggingface.co/HopitAI/moda-fashion-crossdomain) | image β†’ image | 203M | street-to-shop |
121
+ | [moda-fashion-vision-fp16](https://huggingface.co/HopitAI/moda-fashion-vision-fp16) | image β†’ image | 93M vision | edge and mobile, 186 MB |
122
 
123
+ ## How these numbers are measured
 
124
 
125
+ Full corpus, no subsampling β€” every query scored against the entire benchmark corpus
126
+ (42,537–201,624 items). One evaluator for every system, including competitors. Serving
127
+ recipes are selected on held-out development data and frozen in writing before any benchmark
128
+ is scored. Losses are shown, not hidden.
129
+
130
+ Receipts, per-query results and confidence intervals: **[github.com/hopit-ai/Moda](https://github.com/hopit-ai/Moda)**
131
+ Β· Benchmarks: **[hopit-ai.github.io/Moda](https://hopit-ai.github.io/Moda/)**
132
+ Β· Hosted retrieval and agentic search: **[hopit.ai](https://hopit.ai)**