HyperView

Open-source dataset curation + embedding visualization (Euclidean + Poincaré disk)

Try the live demo on HuggingFace Spaces

Features

• Dual-Panel UI: Image grid + scatter plot with bidirectional selection
• Multi-Layout Visualizations: Explore Euclidean, Poincare, and spherical layouts in 2D or 3D with UMAP or PCA projections
• HuggingFace Integration: Load datasets directly from HuggingFace Hub
• Fast Embeddings: Uses EmbedAnything for CLIP-based image embeddings

Updates

• 01-02-26 — The Geometry of Image Embeddings, Hands-on Coding Workshop (Berlin Computer Vision Group)
• 17-01-26 — The Geometry of Image Embeddings, Hands-on Coding Workshop, Part I (Berlin Computer Vision Group)
• 11-12-25 — Hacker Room Demo Day #2 (Merantix AI Campus Berlin) — First version of HyperView presented

Quick Start

Docs: docs/datasets.md · docs/colab.md · CONTRIBUTING.md · TESTS.md

Installation

uv pip install hyperview

Run HyperView

hyperview \
  --dataset cifar10_demo \
  --hf-dataset uoft-cs/cifar10 \
  --split train \
  --image-key img \
  --label-key label \
  --samples 500 \
  --model openai/clip-vit-base-patch32 \
  --layout euclidean \
  --layout poincare

This will:

1. Use dataset cifar10_demo
2. Load up to 500 samples from CIFAR-10
3. Compute CLIP embeddings
4. Generate Euclidean and Poincare visualizations
5. Start the server at http://127.0.0.1:6262

You can also launch with explicit dataset/model/projection args:

hyperview \
  --dataset imagenette_clip \
  --hf-dataset fastai/imagenette \
  --split train \
  --image-key image \
  --label-key label \
  --samples 1000 \
  --model openai/clip-vit-base-patch32 \
  --method umap \
  --layout euclidean

Python API

import hyperview as hv

# Create dataset
dataset = hv.Dataset("my_dataset")

# Load from HuggingFace
dataset.add_from_huggingface(
    "uoft-cs/cifar100",
    split="train",
    max_samples=1000
)

# Or load from local directory
# dataset.add_images_dir("/path/to/images", label_from_folder=True)

# Compute embeddings and visualization
dataset.compute_embeddings(model="openai/clip-vit-base-patch32")
dataset.compute_visualization()

# Launch the UI
hv.launch(dataset)  # Opens http://127.0.0.1:6262

Google Colab

See docs/colab.md for a fast Colab smoke test and notebook-friendly launch behavior.

Why Hyperbolic?

Traditional Euclidean embeddings struggle with hierarchical data. In Euclidean space, volume grows polynomially ($r^d$), causing Representation Collapse where minority classes get crushed together.

Hyperbolic space (Poincaré disk) has exponential volume growth ($e^r$), naturally preserving hierarchical structure and keeping rare classes distinct.

Try the live demo on HuggingFace Spaces→

Community

Weekly Open Discussion — Every Tuesday at 15:00 UTC on Discord

Join us to see the latest features demoed live, walk through new code, and get help with local setup. Whether you're a core maintainer or looking for your first contribution, everyone is welcome.

Contributing

Development setup, frontend hot-reload, and backend API notes live in CONTRIBUTING.md.

Related projects

• hyper-scatter: High-performance WebGL scatterplot engine (Euclidean + Poincaré) used by the frontend: https://github.com/Hyper3Labs/hyper-scatter
• hyper-models: Non-Euclidean model zoo + ONNX exports : https://github.com/Hyper3Labs/hyper-models

License

MIT License - see LICENSE for details.