SAM3 Autolabeling with LightlyStudio

Auto-label your dataset with SAM3's text prompts, then curate with LightlyStudio's embedding view to find and fix systematic errors — in under an hour.

Contents

File	Description
samv3_autolabel_lightlystudio.ipynb	End-to-end notebook: download Oxford Pets, run SAM3, curate in LightlyStudio, export COCO
samv3_blog_post.md	Full written walkthrough of the pipeline
samv3_video_script.md	Video script for the screen-recording demo

Quickstart

# Install dependencies
pip install lightly-studio kaggle transformers torch opencv-python-headless
pip install "git+https://github.com/lightly-ai/lightly-studio-plugins.git#subdirectory=plugins/sam3_segmentation/"

# HuggingFace access required for SAM3
# Request at: https://huggingface.co/facebook/sam3
huggingface-cli login

Then open samv3_autolabel_lightlystudio.ipynb and run all cells.

What you get

• SAM3 text-prompted segmentation — type "cat and dog", get labeled masks back
• LightlyStudio curation — embedding map shows systematic SAM3 errors as clusters
• 93% cost reduction vs. manual labeling
• COCO export — ready for Mask R-CNN, YOLOv8-seg, or any training framework

Requirements

• Python 3.10+
• GPU recommended (CPU works but is slower)
• Kaggle API key for dataset download
• HuggingFace access for SAM3 weights

---

Built by Lightly AI