SAM3 — Segment Anything Model 3

This is an ungated mirror of Meta's SAM3 model weights, redistributed under the SAM License for easier access.

What is SAM3?

SAM3 (Segment Anything Model 3) is Meta AI's state-of-the-art foundation model for promptable segmentation in images and videos. It supports:

• Text prompts — segment objects by describing them (e.g. "the dog")
• Bounding box prompts — precise region targeting
• Point prompts — click-to-segment
• Video tracking — track and segment objects across video frames
• Automatic mask generation — segment all objects in a scene

Files

⚠️ Conversion Note

The sam3.safetensors in this repo is converted directly from the official sam3.pt checkpoint using:

import torch
from safetensors.torch import save_file

state_dict = torch.load("sam3.pt", map_location="cpu")
save_file(state_dict, "sam3.safetensors")

This preserves the original key names (detector.*, tracker.*). Do not use the HuggingFace Transformers-format safetensors from facebook/sam3 — those have different key names (detector_model.*) and are incompatible with build_sam3_video_model().

Usage

# With the official sam3 package
pip install sam3

from sam3.model_builder import build_sam3_image_model
model = build_sam3_image_model(checkpoint_path="sam3.safetensors")

License

This model is distributed under the SAM License — see the LICENSE file. Key points:

• ✅ Commercial use permitted
• ✅ Redistribution permitted (with license included)
• ✅ Derivative works permitted
• ❌ No military/warfare, nuclear, or espionage use
• ❌ No reverse engineering

Credits

• Original model by: Meta AI (FAIR)
• Original HuggingFace repo: facebook/sam3
• Paper: Segment Anything Model 3 (submitted to ICLR 2026)
• Redistributed by: Æmotion Studio for use with ComfyUI-FFMPEGA