GEM v6 (ours)
mAP@20
Size
Size
About This Model
Overview
Nyris General Embedding Model (GEM) v6 is a generic, vision-only embedding model designed for broad instance-level image retrieval across industrial domains, including manufacturing, DIY, and retail. Unlike models tailored to a single vertical, GEM v6 is trained to handle a wide variety of product categories while retaining the ability to distinguish between visually similar instances.
Architecture
GEM v6 uses an MLCD ViT-bigG vision backbone and is evaluated as a vision-only image encoder that produces 768-dimensional global image embeddings (Matryoshka slice at 768 dimensions). The model is engineered for efficient nearest-neighbor retrieval and is used in a pure image-to-image instance retrieval setting without any task-specific adaptation.
Capabilities
GEM v6 is intended for visual search scenarios with heterogeneous capture conditions, including variation in lighting, background clutter, viewpoint, and differences between catalog imagery and user-captured photos. In this benchmark, it serves as a general-purpose vision baseline for image-to-image product retrieval across multiple domains.
Performance Across Datasets
| Dataset | Category | R@1 | R@5 | mAP |
|---|---|---|---|---|
| Stanford Online Products | E-commerce | 87.06% | 94.57% | 73.09% |
| TOPEX | Industrial | 81.37% | 88.72% | 68.32% |
| Products-10K | E-commerce | 77.50% | 91.86% | 66.55% |
| Furniture v1 | Retail | 76.28% | 88.41% | 62.46% |
| Clips-and-Connectors v1 | Industrial | 70.17% | 86.87% | 43.02% |
| DIY v1 | Hardware/DIY | 34.92% | 58.97% | 43.45% |
| Automotive v1 | Automotive | 33.22% | 53.44% | 36.17% |
| Average | 65.79% | 80.41% | 56.15% | |
ILIAS
Showing the results for ILIAS separately, as it is excluded from the average calculations. Because not all models are evaluated on ILIAS.
| Dataset | Category | R@1 | R@5 | mAP@20 | mAP@1000 |
|---|---|---|---|---|---|
| ILIAS | Instance Retrieval | 82.55% | 87.66% | 49.34% | 50.11% |