HYPER3-CLIP V1 · BUILT WITH HAYSTACK
Move the slider. Follow the images.
·
The example is encoded once. Hyper3 moves along the same ray, from near the origin outward, and Haystack retrieves the nearest images in its native hyperbolic space. The query and its direction stay fixed.
CLIP uses the same unchanged query. Its cosine ranking stays fixed as you move the slider.
The slider explores radial retrieval; it does not guarantee a semantic hierarchy. Very close to the origin, unrelated queries can converge on the same images. This view uses Lorentz distance, not the cone score used in the caption-comparison experiment.
All 101 slider positions were computed with the same Haystack pipeline before publication. The browser displays those recorded results, so no server or model download is needed.
42 Creative Commons images. Photo credits and licenses ↗