org card: under-250M class first, then all-systems
Browse files
README.md
CHANGED
|
@@ -28,7 +28,29 @@ short_description: SOTA fashion retrieval, measured and open.
|
|
| 28 |
</section>
|
| 29 |
|
| 30 |
<section>
|
| 31 |
-
<h2>Text-to-image
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
<p>
|
| 33 |
Full-corpus MAP@10 through one shared pipeline, identical preprocessing per
|
| 34 |
model, no gallery subsampling anywhere. Best per row in bold with π₯, second π₯.
|
|
|
|
| 28 |
</section>
|
| 29 |
|
| 30 |
<section>
|
| 31 |
+
<h2>Text-to-image β models under 250M parameters</h2>
|
| 32 |
+
<p>The size class most deployments use. Full-corpus MAP@10; best per row bold π₯, second π₯:</p>
|
| 33 |
+
<table>
|
| 34 |
+
<tr><th>benchmark (corpus)</th><th>FashionSigLIP 203M</th><th>MODA 203M</th><th>MODA Pro Lite 213M</th></tr>
|
| 35 |
+
<tr><td>KAGL (44K)</td><td>0.2769</td><td>0.2890 π₯</td><td><b>0.3185</b> π₯</td></tr>
|
| 36 |
+
<tr><td>Polyvore (94K)</td><td>0.3665</td><td>0.3726 π₯</td><td><b>0.3997</b> π₯</td></tr>
|
| 37 |
+
<tr><td>Atlas (78K)</td><td>0.1826</td><td>0.1884 π₯</td><td><b>0.1945</b> π₯</td></tr>
|
| 38 |
+
<tr><td>Fashion200K (202K)</td><td>0.1858 π₯</td><td><b>0.1947</b> π₯</td><td>0.1802</td></tr>
|
| 39 |
+
<tr><td>DeepFashion In-Shop (53K)</td><td>0.1587 π₯</td><td><b>0.1703</b> π₯</td><td>0.1031</td></tr>
|
| 40 |
+
<tr><td>DeepFashion Multimodal (43K)</td><td><b>0.0148</b> π₯</td><td>0.0147 π₯</td><td>0.0118</td></tr>
|
| 41 |
+
</table>
|
| 42 |
+
<p>
|
| 43 |
+
Every benchmark in this class is led by a MODA-family model except one, where
|
| 44 |
+
frozen FashionSigLIP keeps a 0.8% edge.
|
| 45 |
+
<b><a href="https://huggingface.co/HopitAI/moda-pro-lite">MODA Pro Lite</a> owns catalog
|
| 46 |
+
and title search</b> (KAGL +10.2%, Polyvore +7.3%, Atlas +3.2% over MODA) from one
|
| 47 |
+
checkpoint, no serving recipe; <b>MODA owns captions and instance retrieval</b>
|
| 48 |
+
(4 of 6 wins over FashionSigLIP significant under paired bootstrap).
|
| 49 |
+
</p>
|
| 50 |
+
</section>
|
| 51 |
+
|
| 52 |
+
<section>
|
| 53 |
+
<h2>Text-to-image β all systems, including larger models</h2>
|
| 54 |
<p>
|
| 55 |
Full-corpus MAP@10 through one shared pipeline, identical preprocessing per
|
| 56 |
model, no gallery subsampling anywhere. Best per row in bold with π₯, second π₯.
|