SAM 2

SAM 2

SAM 2 (Segment Anything Model 2) from Meta FAIR extends the original SAM to videos, enabling real-time promptable object segmentation that propagates across frames with a streaming memory architecture. It achieves state-of-the-art results on 17 diverse video segmentation benchmarks.

Key Features

• Video segmentation: track any object through a video clip given a single click or bounding box prompt on any frame
• Image segmentation: full backward-compatible drop-in for SAM 1 with improved accuracy and 6× faster inference
• Streaming memory: efficient memory encoder caches past frames so the model can handle long videos without quadratic cost
• Multi-prompt: combine point, box, and mask prompts; add corrective prompts mid-video
• Pretrained checkpoints: SAM 2 Hiera-T/S/B+/L — trade accuracy for speed
• Data engine: 50.9M mask annotations on 642.6K videos used for training, released as SA-V dataset

Quick Start

pip install samv2

import torch
from sam2.build_sam import build_sam2
from sam2.sam2_image_predictor import SAM2ImagePredictor

checkpoint = "checkpoints/sam2.1_hiera_large.pt"
model_cfg = "configs/sam2.1/sam2.1_hiera_l.yaml"
predictor = SAM2ImagePredictor(build_sam2(model_cfg, checkpoint))

with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
    predictor.set_image(image)  # numpy HxWx3
    masks, scores, _ = predictor.predict(point_coords=[[500, 375]], point_labels=[1])

npx ai-supply add sam2-segment-anything-2

Curated mirror of the open-source SAM 2 (Apache-2.0). Get it from the source.