| --- |
| license: other |
| extra_gated_fields: |
| First Name: text |
| Last Name: text |
| Date of birth: date_picker |
| Country: country |
| Affiliation: text |
| Job title: |
| type: select |
| options: |
| - Student |
| - Research Graduate |
| - AI researcher |
| - AI developer/engineer |
| - Reporter |
| - Other |
| geo: ip_location |
| By clicking Submit below I accept the terms of the license and acknowledge that the information I provide will be collected stored processed and shared in accordance with the Meta Privacy Policy: checkbox |
| extra_gated_description: >- |
| The information you provide will be collected, stored, processed and shared in |
| accordance with the [Meta Privacy Policy](https://www.facebook.com/privacy/policy/). |
| extra_gated_button_content: Submit |
| language: |
| - en |
| pipeline_tag: mask-generation |
| library_name: transformers |
| tags: |
| - sam3 |
| --- |
| |
| # SAM 3 |
|
|
| This repository mirrors the official **Segment Anything Model 3 (SAM 3)** weights released by Meta Superintelligence Labs. SAM 3 is a unified foundation model for prompt-driven segmentation in images and videos. It supports open-vocabulary text prompts and visual prompts (points/boxes/masks). Compared to SAM 2, SAM 3 exhaustively segments each instance of a requested concept and reaches ~75–80% of human-level performance on the SA-CO benchmark (270K unique concepts). |
|
|
| ## Highlights |
| - **Presence token** improves discrimination between closely related prompts. |
| - **Decoupled detector + tracker** scales better for long video sequences. |
| - **4M+ automatically annotated concepts** ensure broad coverage of open-world categories. |
|
|
| > Original paper: *SAM 3: Segment Anything with Concepts* (Meta AI, 2024). |
| > Resources: [Project Page](https://ai.meta.com/sam3) · [Demo](https://segment-anything.com/) |
|
|
| ## Files Included |
| - `sam3.safetensors` — detector and tracker weights for image + video segmentation. |
| - Tokenizer/config assets should be copied from the official `facebookresearch/sam3` repository; this mirror only repackages the safetensors weights for self-hosting. |
|
|
| ## Quickstart |
| ```bash |
| pip install torch==2.7.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126 |
| pip install git+https://github.com/facebookresearch/sam3.git |
| |
| python - <<'PY' |
| from sam3 import build_sam3_image_model |
| from sam3.model.sam3_image_processor import Sam3Processor |
| model = build_sam3_image_model( |
| bpe_path="sam3/assets/bpe_simple_vocab_16e6.txt.gz", |
| device="cuda", |
| eval_mode=True, |
| checkpoint_path="sam3.safetensors", |
| load_from_HF=False, |
| ) |
| processor = Sam3Processor(model, device="cuda") |
| state = processor.set_image("your_image.jpg") |
| state = processor.set_text_prompt("white bicycle", state) |
| print(state["masks"].shape) |
| PY |
| ``` |
|
|
| ## Integration Notes |
| These mirrored weights are used in the **AILab SAM3 ComfyUI node** (RMBG edition) to enable promptable segmentation workflows directly inside ComfyUI. The node loads `sam3.safetensors`, tokenizer assets, and the SAM3 processors locally, so the entire pipeline stays compatible even when offline. |
|
|
| ## License & Usage |
| - This mirror preserves Meta's original weights and is subject to the license on [facebook/sam3](https://huggingface.co/facebook/sam3). You must accept Meta's terms before downloading the official release. |
| - When hosting this file in your own Hugging Face repository, keep this notice and credit the original authors. |
| - Cite the SAM 3 paper for any research or product that builds upon these weights. |
|
|
|
|
