Note

This is a community plugin, an external project maintained by its respective author. Community plugins are not part of FiftyOne core and may change independently. Please review each plugin’s documentation and license before use.

GitHub Repo

FiftyComfy#

img

A visual, node-based workflow editor for FiftyOne – powered by LiteGraph.js (the same engine behind ComfyUI). Build dataset curation, analysis, and model evaluation pipelines by connecting nodes on a canvas, without writing code.

Installation#

fiftyone plugins download https://github.com/harpreetsahota204/FiftyComfy --overwrite

Getting Started#

  1. Launch the FiftyOne App with a dataset loaded:

import fiftyone as fo
import fiftyone.zoo as foz

dataset = foz.load_zoo_dataset("quickstart")
session = fo.launch_app(dataset)

  1. Open the FiftyComfy panel from the + button in the App’s panel selector.

  2. Right-click on the canvas (or double-click) to open the node search menu and start adding nodes.

  3. Connect nodes by dragging from an output slot to an input slot. All nodes pass FO_VIEW (a FiftyOne DatasetView) through connections.

  4. Click Run Workflow to execute the graph. Results are pushed back into the FiftyOne App.

How It Works#

FiftyComfy is a hybrid FiftyOne plugin:

  • Frontend: A LiteGraph.js canvas wrapped in a React component, providing the ComfyUI-style node editor with pan, zoom, context menus, connection validation, and widget rendering.

  • Backend: Python operators that receive the serialized graph, topologically sort it, and execute each node against the current dataset using the FiftyOne SDK.

When you click “Run Workflow”, the graph is serialized to JSON and sent to the Python backend. The executor walks the graph in topological order, passing FiftyOne DatasetView objects between nodes. Output nodes like “Set App View” push the final view into the FiftyOne App grid.

Nodes#

Source (orange)#

Entry points for your workflow. Every graph needs at least one source node.

Node

Description

Current Dataset

Uses the dataset currently loaded in the App. Displays the dataset name on the node.

Load Saved View

Loads a named saved view from the dataset. Dropdown populated from your dataset’s saved views.

View Stages (blue)#

Filter, sort, and transform your data. These correspond to FiftyOne view stages.

Node

Description

Widgets

Match

Filter samples by a ViewField expression

expression (text) – e.g., F('confidence') > 0.5

Filter Labels

Filter detections/classifications within a label field

field (dropdown), expression (text), only_matches (toggle)

Match Labels

Match samples containing labels satisfying a condition

field (dropdown), filter (text), bool (toggle)

Sort By

Sort samples by a field

field (dropdown), reverse (toggle)

Sort By Similarity

Sort by similarity to a reference (requires a similarity index)

brain_key (text), k (number), reverse (toggle)

Limit

Limit number of samples

count (number)

Match Tags

Filter samples by tags

tags (text, comma-separated)

Take

Randomly sample N samples

count (number), seed (number)

Shuffle

Randomly shuffle samples

seed (number)

To Patches

Convert to a patches view from a label list field (Detections, Polylines, Keypoints)

field (dropdown)

Map Labels

Remap label values using a mapping dict

field (dropdown), map (text) – e.g., {"cat": "animal"}

Brain (purple)#

FiftyOne Brain methods for dataset analysis and curation.

Node

Description

Key Widgets

Compute Embeddings

Compute embeddings using a zoo model

model (dropdown), embeddings_field (text)

Compute Visualization

Compute UMAP/t-SNE/PCA visualization

brain_key, method (dropdown), num_dims, embeddings

Compute Similarity

Create a similarity index

brain_key, backend (dropdown), embeddings

Compute Uniqueness

Score sample uniqueness

uniqueness_field, embeddings

Compute Representativeness

Score how representative each sample is

representativeness_field, method (dropdown), embeddings

Compute Mistakenness

Estimate annotation mistake likelihood

pred_field (dropdown), label_field (dropdown)

Compute Hardness

Compute sample hardness

predictions_field (dropdown)

Find Exact Duplicates

Find identical media files

(no widgets)

Find Near Duplicates

Find near-duplicate samples

threshold (number), embeddings

Detect Leaky Splits

Find data leaks across splits

splits (text), threshold (number)

Manage Brain Run

Rename or delete a brain run

brain_key (dropdown), action (dropdown), new_name (text)

Model (green)#

Apply models from the FiftyOne Model Zoo.

Node

Description

Key Widgets

Apply Zoo Model

Run inference with a zoo model

model (dropdown – 17 popular models), label_field (text), confidence_thresh (number), store_logits (toggle)

Evaluation (green)#

Evaluate model predictions against ground truth.

Node

Description

Key Widgets

Evaluate Detections

Evaluate detection predictions (COCO or Open Images style)

pred_field (dropdown), gt_field (dropdown), eval_key (text), method (dropdown)

Evaluate Classifications

Evaluate classification predictions

pred_field (dropdown), gt_field (dropdown), eval_key (text)

Manage Evaluation

Rename or delete an evaluation run

eval_key (dropdown), action (dropdown), new_name (text)

Output (yellow)#

Terminal nodes that push results back to the FiftyOne App.

Node

Description

Widgets

Set App View

Push the resulting view into the App grid

(no widgets)

Save View

Save the view as a named saved view

name (text), description (text), overwrite (toggle)

Dynamic Dropdowns#

Field dropdowns are automatically populated from your dataset’s schema when the panel loads:

  • Label fields (Detections, Classifications, Polylines, Keypoints) are used by Filter Labels, Match Labels, To Patches, Map Labels, and evaluation nodes.

  • Patches fields (Detections, Polylines, Keypoints) are used by To Patches.

  • All fields are used by Sort By.

  • Saved views, brain runs, and evaluation runs populate their respective management dropdowns.

  • Label classes (distinct values like ["car", "person", "dog"]) are shown as a preview at the bottom of Filter Labels and Match Labels nodes.

Example Workflows#

Filter and view high-confidence detections#

Current Dataset --> Filter Labels (field: predictions, expr: F('confidence') > 0.8) --> Set App View

Compute embeddings, then visualize#

Current Dataset --> Compute Embeddings (model: clip-vit-base32-torch) --> Compute Visualization (method: umap) --> Set App View

Apply a model and evaluate#

Current Dataset --> Apply Zoo Model (model: faster-rcnn-resnet50-fpn-coco-torch, label_field: predictions) --> Evaluate Detections (pred: predictions, gt: ground_truth) --> Sort By (field: eval_fp, reverse: true) --> Set App View

Find near duplicates#

Current Dataset --> Compute Embeddings --> Find Near Duplicates (threshold: 0.1) --> Set App View

Saving and Loading Workflows#

  • Save: Click the “Save” button in the toolbar and enter a name. Workflows are stored in browser localStorage.

  • Load: Click “Load” to see saved workflows and click one to restore it.

  • Clear: Click “Clear” to reset the canvas.

Canvas Controls#

All standard LiteGraph controls work out of the box:

  • Pan: Middle-click drag or Ctrl+click drag

  • Zoom: Mouse wheel

  • Add node: Right-click for context menu, or double-click for search

  • Connect: Drag from an output slot to an input slot

  • Delete: Select node(s) and press Delete/Backspace

  • Copy/Paste: Ctrl+C / Ctrl+V

  • Multi-select: Click and drag to box-select

  • Collapse node: Double-click the title bar

License#

Apache 2.0