HYPER3-CLIP V1 · BUILT WITH HAYSTACK
Move from a broader scene to a closer match.
·
Each scene searches a small, fixed collection of full photographs. Move the slider to shift one query outward along the same direction; the three nearest images appear first. Both models search the identical collection.
The figure shows the query’s native radial movement and model cone. Each scene is rotated for layout, so screen angles are not comparable across scenes. Image-to-query distances are preserved; image-to-image distances are not.
The cone illustrates the model’s geometry; it does not filter the results. CLIP uses the same description and keeps its cosine-similarity ranking as the slider moves.
The Haystack pipeline computed all slider positions beforehand; the browser displays those results.
Selected examples illustrate traversal, not a quality benchmark. Photo credits ↗
hyper³labs