Interactive geometric workbench for embedding space analysis.
HyperView turns image datasets into live embedding workspaces. Load samples, compute embeddings, inspect Euclidean, hyperbolic, or spherical layouts, select clusters and outliers, and keep media, labels, layouts, selections, panels, and tools in one local workspace.
Try the live demo on Hugging Face Spaces
uv tool install --upgrade hyperview
hyperview skill installFor a project-local GitHub Copilot skill:
hyperview skill install --scope project --agent github-copilot --yesimport hyperview as hv
dataset = hv.Dataset("cifar100")
dataset.add_from_huggingface("uoft-cs/cifar100", split="train", max_samples=1000)
dataset.compute_embeddings(model="openai/clip-vit-base-patch32")
dataset.compute_visualization()
hv.launch(dataset)hyperview workspace create imagenette-demo \
--dataset imagenette_clip_20260411 \
--activate
hyperview serve \
--workspace imagenette-demo \
--dataset imagenette_clip_20260411Agents and scripts can control the running workspace:
hyperview ui layout set --workspace imagenette-demo --layout-key <layout-key>
hyperview ui selection set --workspace imagenette-demo --ids sample-1,sample-8
hyperview extension add .hyperview/extensions/selection-profile --workspace imagenette-demo
hyperview tools run selection_profile.summarize --workspace imagenette-demo- Ingests image data from Hugging Face datasets or local folders.
- Computes embeddings with built-in or custom providers.
- Projects spaces into Euclidean, Poincare/hyperbolic, and spherical layouts.
- Links scatter points to thumbnails, labels, selections, and nearest neighbors.
- Adds dataset-specific Python tools and native frontend panels.
- Gives coding agents a CLI control plane for the same workspace humans inspect.
Embedding failures often hide in the shape of the space: collapsed classes, weak separation, hierarchy, long-tail samples, and boundary cases. HyperView lets you inspect those structures through multiple geometric views instead of a single fixed projection.
- hyper-scatter: WebGL scatterplot engine for Euclidean and Poincare views.
- hyper-models: Non-Euclidean model zoo and ONNX exports.
- hyper-lrp: Attribution tools for inspecting model evidence.
Join the Hyper3Labs Discord for demos, setup help, and project discussion.
MIT License. See LICENSE.
If you use HyperView in research, please cite:
@software{hyperview2026,
author = {{Hyper3Labs}},
title = {HyperView: An Interactive Geometric Workbench for Embedding Space Analysis},
year = {2026},
version = {0.5.0},
url = {https://github.com/Hyper3Labs/HyperView}
}