SAM3 Auto-Labeling

Auto-label thousands of images in minutes using Meta's Segment Anything 3 with plain-text prompts. 3LC captures image and instance embeddings during prediction — explore the embedding space in the Dashboard to catch mislabeled samples, strip out near-duplicate frames that would hurt model generalization, and remove low-confidence predictions before they pollute your training set. Accept only the right data with just a touch of human-in-the-loop review — train better models with less data and zero surprises.

Auto-label

SAM3 predicts masks and bounding boxes for every image from your text prompts alone

Inspect embeddings

Explore image and instance embedding space to verify quality and catch outliers

Cut redundancy

Strip duplicates and near-similar frames — keep only the data that improves the model

Train smarter

Zero-weight start — accept what you trust, skip the rest, get the best model with the least data

1 Setup & preview

2 Create table & run prediction

Config

1 Setup & Preview

Config Name

Image Source

Folder Path

Table URL

Point to an image folder or existing 3LC table. SAM3 will process every image in the source.

Labels (text prompts)

Define the object classes you want to detect. Use descriptive text prompts that SAM3 can understand (e.g. "fish", "person walking").

Modality

Confidence: 0.20

Preview

Add labels and click
Preview Random Image

2 Create Table & Run Prediction

Creates a 3LC table from the images, then automatically runs SAM3 prediction on all samples with UMAP embeddings.

3LC Project

Dataset Name

Table Name

Run Name

Max Images

Embeddings

Device

Progress

No active job.

Log