Hello World
The same 300 iNaturalist photos mapped in Euclidean and spherical space by CLIP and in hyperbolic space by Hyper3-CLIP.
iNaturalist images · three geometries · custom panel
HyperView Spaces
Explore model comparisons through prepared examples in interactive HyperView workspaces, with nothing to install. Most compare Hyper3-CLIP, our hyperbolic image–text model, with OpenAI CLIP. Each links to its source code.
ABO Catalog
Find the right Amazon Berkeley Objects product from a photo or a detailed request.
Images · embeddings · ranked neighbours
Fashion Products
Find the same DeepFashion garment from another photo or a written request.
Product photos · text · ranked matches
Precision Regions
Find the one image region that matches a written description, on RefCOCOg scenes.
Scenes · crops · ranked regions
Logo Search
Match detailed creative briefs to the right logo in a 160-logo catalog.
Logos · creative briefs · style map
GeoSpatial
Check whether an aerial tile's nearest neighbours share its land use, across 60 NWPU-RESISC45 tiles.
Aerial imagery · neighbours · archive maps
Visual Safety
Start from one confirmed item and compare which Open Images photos each model queues for review.
Images · review batches · ranking quality
Jaguar Multi-Geometry
Euclidean, spherical and Poincaré views of jaguar re-identification embeddings, with nearest neighbours from the original study.
Jaguar imagery · re-identification · multiple geometries
Build your own
Copy a folder from hyperview-spaces, set your dataset and models at the top of demo.py, and publish with hyperview publish. See the HyperView docs.