# Visualize FiftyOne Embeddings

Image datasets can hide clusters, outliers, and label problems. This skill guides Claude, Codex, and Claude Code through FiftyOne embedding visualizations.

## Install

```bash
npx skillstore add adonaivera/fiftyone-embeddings-visualization
```

## Metadata

- Status: approved
- Slug: adonaivera-fiftyone-embeddings-visualization
- Skillstore revision: r1
- Version status: missing
- Tree hash: dbaae6d76d79338d6b8a55c57aea71e281ebe54a5c35d85807112b9bf16f3d32
- Author: AdonaiVera
- GitHub username: AdonaiVera
- License: MIT
- Repository: https://github.com/AdonaiVera/fiftyone-skills/tree/main/embeddings-visualization/skills/fiftyone-embeddings-visualization
- Ref: 6eec4a59f63ff83440a84c18d1aef79743bc714e
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: allowed
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands, network
- Quality score: 77
- Quality tier: bronze
- Public page: https://skillstore.pages.dev/skills/adonaivera-fiftyone-embeddings-visualization
- Manifest: https://skillstore.pages.dev/api/skills/adonaivera-fiftyone-embeddings-visualization/manifest

## Capabilities

- Sets the active FiftyOne dataset context before running visualization steps.
- Launches the FiftyOne App so Brain plugin operators can execute.
- Discovers Brain operators and schemas before selecting parameters.
- Computes or reuses image embeddings for similarity and visualization workflows.
- Creates 2D UMAP, t-SNE, or PCA visualizations for the Embeddings panel.
- Guides users through coloring, filtering, and inspecting points in the app.

## Use Cases

- Explore Dataset Structure: Visualize image clusters, review dense regions, and find isolated samples before labeling or quality review.
- Diagnose Model Errors: Compare ground truth and prediction patterns in embedding space to locate confusing classes and likely mistakes.
- Review Representation Quality: Compare CLIP, DINOv2, ResNet, or MobileNet embeddings with UMAP, t-SNE, or PCA views.

## Prompt Templates

### Visualize a Dataset

```
Use the FiftyOne embeddings visualization workflow for dataset my-dataset. Check for existing embeddings, create a 2D UMAP view if needed, and tell me how to open it in the app.
```

### Find Outliers

```
Analyze dataset my-dataset for unusual images. Compute embeddings if needed, create an outlier-focused visualization, and explain where I should inspect isolated samples.
```

### Compare Classes

```
Create an embedding visualization for my-dataset and guide me through coloring points by ground truth and predictions. Highlight signs of class overlap or label issues.
```

### Compare Embedding Models

```
For my-dataset, compare CLIP and DINOv2 embedding visualizations. Reuse existing fields when available, choose suitable reduction methods, and summarize the visual differences.
```

## Limitations

- Requires a running FiftyOne MCP server and an image dataset already loaded.
- Requires the FiftyOne Brain plugin for delegated visualization operators.
- Large datasets can take minutes or hours to embed depending on model and hardware.
- UMAP requires the separate umap-learn package, or users must choose t-SNE or PCA.

## Best Practices

- Discover available operators and schemas before choosing parameters.
- Reuse existing embedding fields when possible to avoid repeated computation.
- Start with a subset for very large datasets before processing every sample.

## Anti Patterns

- Do not filter samples by brain key when opening the Embeddings panel.
- Do not assume UMAP is available when umap-learn is not installed.
- Do not compute new embeddings when a suitable field already exists.

## Security Audit

- Audited at: 2026-07-05T01:36:46.79\+00:00
- Summary: All static command findings are Markdown code fences, inline parameter names, or FiftyOne MCP examples rather than executable Ruby or shell backticks. The network findings are localhost app directions and public documentation links, and no prompt injection, exfiltration intent, or malicious behavior was found.

## Stats

- Views: 320
- Downloads: 11
- Favorites: 0
- Popularity score: 0