Overview

hyper-models is a catalog of non-Euclidean embedding models. Each model has a short name, and load() returns an object that encodes images the same way whichever runtime sits underneath.

Install

uv pip install hyper-models

The base install has no PyTorch dependency and runs the ONNX models, HyCoCLIP and MERU. Models that need PyTorch, such as Hyper3-CLIP and UNCHA, need the ml extra:

uv pip install "hyper-models[ml]"

Load and encode

import hyper_models
from PIL import Image

model = hyper_models.load("hycoclip-vit-s")
model.geometry   # 'hyperboloid'
model.dim        # 513

embeddings = model.encode_images([Image.open("image.jpg")])   # shape (1, 513)

Weights download from the Hugging Face Hub on first use and are cached.

load() also takes revision, token, local_files_only, device and local_path. Pin revision when new embeddings must match an index you built earlier.

Text and images in one space

Hyper3-CLIP encodes text and images into the same 513-coordinate Lorentz space:

model = hyper_models.load("hyper3-clip-v1")
images = model.encode_images([Image.open("sofa.jpg")])
texts = model.encode_texts(["a grey velvet sofa"])

hyper3-clip-v1 is gated. Accept its terms on its Hugging Face page and run hf auth login before the first download.

Inspect the catalog

hyper_models.list_models()
info = hyper_models.get_model_info("hycoclip-vit-s")
info.hub_id    # 'mnm-matin/hyperbolic-clip'
info.loader    # 'onnx'
info.license   # 'CC-BY-NC'

Check each model's licence before commercial use. The full list is on Models.