---
title: latent-inspector
emoji: 🔬
colorFrom: blue
colorTo: indigo
sdk: static
app_file: index.html
pinned: true
license: mit
short_description: How self-supervised vision models carve the same image.
tags:
  - vision
  - self-supervised
  - representation-analysis
  - cka
  - interpretability
---

# latent-inspector — public front door

Browser-first overview of [`latent-inspector`](https://github.com/AbdelStark/latent-inspector), a Rust CLI that probes how different self-supervised vision encoders represent the same image.

This Space is the human-readable surface. Two entry points:

- **Presentation** (`slides.html`) — reveal.js deck: *How AI Models See the World*. Walks through the project, the four-model comparison setup, the corrected EUPE and V-JEPA 2 interpretations, and the thesis that training objective reshapes representation geometry.
- **Sample report** (`reports/20260408-123006/report.html`) — self-contained HTML report generated by `latent-inspector compare` for **DINOv2**, **I-JEPA**, **V-JEPA 2**, and **EUPE** on the same elephant image. Per-model metrics, pairwise CKA and k-NN overlap, PCA projections, and exported artifact metadata.

## Companion artifacts on this Hub

| | |
|---|---|
| ONNX model exports (parity-validated) | [Latent Inspector — ONNX Vision Encoders](https://huggingface.co/collections/abdelstark/latent-inspector-onnx-vision-encoders-69e64753d939c08fb92a71f2) |
| Fingerprints dataset | [abdelstark/latent-inspector-fingerprints](https://huggingface.co/datasets/abdelstark/latent-inspector-fingerprints) |
| CLI source | [github.com/AbdelStark/latent-inspector](https://github.com/AbdelStark/latent-inspector) |

## What you can do with the CLI

```bash
cargo install --git https://github.com/AbdelStark/latent-inspector latent-inspector

latent-inspector inspect photo.jpg --model vjepa2-vitl-img16-256
latent-inspector compare photo.jpg --models dinov2-vit-l14,ijepa-vit-h14,vjepa2-vitl-img16-256,eupe-vit-b16
```

The models are auto-downloaded from this Hub on first use. The `compare` command generates a self-contained HTML report like the one linked above.

## License

MIT. Underlying model weights keep their upstream licenses.