Meridian is a vision-language representation model built directly on top of standard CLIP (ViT-B/16) backbones. Unlike typical contrastive models that project image and text vectors into a flat, Euclidean vector space, Meridian maps multimodal features onto a continuous hyperbolic manifold (Lorentz manifold).
nlp machine-learning computer-vision representation-learning clip hyperbolic-geometry hyperbolic multimodal-retrieval meru hyperbolic-space retrival hyperbolic-embeddings contrastive-learning image-text-retrieval vision-language-model vision-language-models multimodal-retrieval-augmented-generation hyperbolic-learning lorentz-manifold
-
Updated
Jun 22, 2026 - Python