hyper3-clip is an image-text embedding model built for hierarchy, variants, and compositional visual queries, ranking the exact visual match general models often miss.
Start with operational region search where the target is only part of the image, then move into logo, catalog, fashion, geospatial, and safety workspaces powered by HyperView.
Compare model accuracy when searching for a small region within a complex scene.
#1A grey couch with pillows on it.
Target
#2Brown chair or love seat facing the tv and shelves.
#3a beige side chair with a brown pillow
#4The couch on the bottom left corner
#5A tan chair between two green tables.
#1Brown chair or love seat facing the tv and shelves.
#2The couch on the bottom left corner
#3Chair closest to curtains
#4A tan chair between two green tables.
#5A person in brown pants sitting on a bench next to a child.
Embedding models like OpenAI CLIP project hierarchical data onto flat Euclidean vectors. Because volume in flat spaces grows only polynomially ($r^d$), leaf categories crowd together. This triggers severe retrieval failures where sibling and cousin variants bleed into each other.
Leaf nodes crowd tightly near the bottom. Sibling details overlap and cause retrieval confusion.
Poincaré Disk naturally expands outward. Branch paths separate cleanly, maintaining distance.
We combine specialized vision-language models with an open-source inspection workbench, then test them against concrete retrieval failure modes by industry.
A specialized vision-language model trained with hierarchy-aware geometry to reduce category and sibling bleed in dense visual catalogs.
An open-source viewer for mapping embedding projections, tracing nearest neighbors, and identifying the failure modes behind bad retrieval results.
| Industry / Dataset | Benchmark | hyper3-CLIP | OpenAI-CLIP | Readout |
|---|---|---|---|---|
| Ecommerce Catalog Retrieval – Amazon Berkeley Objects | ||||
| Retail catalogs500 product images, 20 product types | Product-type mAP | 0.582 | 0.552 | +3.05 pts |
| Retail catalogs50 parsed catalog departments | Department mAP | 0.264 | 0.212 | +5.20 pts |
| Retail catalogsParent category retrieves diverse children | Child coverage@50 | 0.780 | 0.655 | +12.50 pts |
| Fashion Product Search – DeepFashion In-Shop | ||||
| Apparel retailSame-item product image retrieval | mAP | 0.407 | 0.240 | +16.7 pts |
| Apparel retailSame-item first-result recovery | Recall@1 | 0.595 | 0.375 | +22.0 pts |
| Apparel retailSpecific typed product search | Hit@10 | 0.572 | 0.550 | +2.2 pts |
| General Visual Hierarchy – COCO Objects | ||||
| Object search5,000 COCO val images, 80 categories | Category mAP | 0.554 | 0.532 | +2.22 pts |
| Object search12 object supercategories | Supercategory mAP | 0.536 | 0.516 | +2.08 pts |
| Object searchCoverage of child types under broad labels | Child coverage@100 | 0.887 | 0.951 | CLIP +6.40 pts |
Understand how our frontier specialized models are trained from scratch, and explore our latest preprints and technical deep-dives.
Why mainstream ML's reliance on Euclidean manifolds is a mistake, and how hyperbolic spaces can efficiently encode hierarchical data with fewer dimensions.
read post →Send a representative retrieval slice. We compare hyper3-clip with your current baseline and return metrics, ranked examples, and a pilot recommendation.
Provide sample images, queries, expected matches, and your current baseline.
We compare hyper3-clip with your baseline and inspect the missed exact matches.
Receive metrics, ranked examples, and a clear recommendation on whether the model is a fit.